← All posts

AI agent spending limits: how Copilot, Replit, Claude and ChatGPT cap costs, and how to set yours

AI agent spending limits: how to stop AI agents overspending. Vestra blog banner

An AI agent spending limit is a cap on how much an AI agent can spend before it stops and asks a person: per task, per day, per month, per user or per agent. Microsoft Copilot, Replit, Claude and ChatGPT all now let admins set caps, but each one caps its own agents, for its own users, inside its own bill.

This guide explains every kind of cap in plain words, shows exactly how the big AI tools cap costs today, and gives you a simple set of rules to set your own. Every competitor fact comes from the maker's own pages (Microsoft, Replit, Anthropic, OpenAI, Tasklet, Lindy, Cognition, Perplexity, GitHub, Cursor, Google, Stripe, Visa and Mastercard), checked on 22 November 2026. Prices and settings change often, so check the maker's page before you rely on one.

Quick answer

You stop AI agents from overspending with three things: a cap on any single task, a cap on the whole day or month, and a rule that anything over the cap waits for a person's yes. Most tools now give you one or two of these, for their own agents only. Microsoft Copilot caps Cowork and Autopilot per tenant, group and user each month. Replit Missions sets a budget in dollars for each mission. Claude lets owners set organisation, group and member spend limits. ChatGPT Business and Enterprise let admins set per-user credit limits and workspace overage limits. The catch is that each cap lives inside one tool. If your business uses five agents from five makers, you have five budgets, five alert emails and five bills, and nothing that caps the team as a whole. Vestra puts every agent, on every engine, under one per-task cap, one daily cap, one rule for money and one receipt.

Key takeaways

  • A spending limit is a cap that stops an agent, or asks a person, before it spends more than you allowed. An alert only tells you. A cap stops.
  • There are eight common kinds of cap: per task, per day, per month, per user, per group, per agent, per workspace and per purchase.
  • Microsoft's own docs say Cowork credit limits are "a spending safeguard, not a real-time access gate". Users may start tasks after the limit before it takes effect.
  • Replit Missions lets you cap a whole mission in dollars and choose which actions need your permission, set up front.
  • Claude Team and Enterprise owners can set organisation-wide and member spend limits. Anthropic says it's possible to slightly exceed them.
  • ChatGPT Enterprise starts its workspace overage limit at "No limit". Admins have to set one.
  • Tasklet's own guide says card purchases rely on the agent following instructions, with "no enforced per-purchase approval gate or spending limit".
  • Almost every cap today is per person or per tool. Very few tools cap a single task, and none of them caps a mixed team of agents from different makers.
  • The rules that work: cap the task, cap the day, cap each agent's card, require a yes for anything new, and get a receipt for every piece of work.
  • In Vestra, per-task and daily caps, your yes on money and an itemised receipt per credit are live today, the same on every engine.

Contents

  • What is an AI agent spending limit?
  • Why do AI agents need spending limits now?
  • How can an AI agent spend your money?
  • What kinds of spending caps are there?
  • How does Microsoft Copilot cap agent costs?
  • How does Replit cap Missions and Agent spending?
  • How does Claude cap spending?
  • How does ChatGPT cap spending?
  • How do other AI agents cap spending?
  • AI agent spending limits compared, tool by tool
  • What do today's caps miss?
  • How do payment cards and rails cap agent purchases?
  • How do you set spending limits for your AI agents?
  • What should always wait for a person's yes?
  • How much should an AI agent be allowed to spend?
  • What happens when an agent hits its cap?
  • How does Vestra cap what its AI team spends?
  • A week of capped spending in Vestra
  • Common mistakes with AI agent budgets
  • Frequently asked questions
  • Glossary
  • Sources

What is an AI agent spending limit?

An AI agent spending limit is a rule that caps how much an agent may spend, in money or credits, before it stops or asks a person. It can apply to one task, one day, one month, one user, one agent or one purchase. A real limit stops the spending. An alert only warns.

That difference matters more than anything else in this guide. Many tools offer both, and the names are easy to mix up. Here are the words you'll see.

  • Spending limit or spend limit. A cap on total spend for a period. When it's reached, the agent stops or the extra spending waits for a yes.
  • Hard limit. A cap the system enforces. The work stops at the line, or very close to it.
  • Alert or threshold. A message sent when spend reaches a level, such as 75% or 90%. Nothing stops.
  • Budget. Usually a planned amount for a period. Some tools enforce budgets as hard limits. Others only track spend against them.
  • Credits. The unit many AI tools bill in. Microsoft charges $0.01 per Copilot Credit. In Vestra, one credit is one cent.
  • Per-task cap. A limit on what one piece of work may cost. The rarest kind of cap, and the most useful.
  • Approval. A person's yes before a specific action, such as a payment, goes ahead.

A spending limit is not the same as a usage limit. A usage limit, like Claude's five-hour window or Codex's weekly limit, controls how much you can use a plan you already paid for. A spending limit controls how much extra you can be charged. Some tools blur the two, because once you run out of included usage, the tool can keep going on paid credits.

Why do AI agents need spending limits now?

AI agents need spending limits now because they work on their own, for longer, and many are billed by use rather than by seat. An agent that runs for hours, retries, calls tools and buys things can run up a surprise bill. In the past few weeks, Microsoft, Replit, Anthropic and OpenAI all added new cost controls.

A chat assistant used to cost a flat monthly fee. You typed, it answered, and the price didn't change if you asked a hard question. Agents broke that model. They do real work: they browse, write code, run tests, send email and call other tools. Each step uses model time and compute. Some steps cost real money outside the AI tool, like an ad test or a software plan.

So the big makers moved to usage-based billing, and with it came caps. Look at the past few weeks alone:

  • Microsoft Ignite (17 November). Microsoft said Agent 365 adds cost management, letting admins "apply spending limits and alerts at the user, group, organization, or resource level". Microsoft's Book of News lists it as a preview. Copilot Cowork, generally available since June, already has spending limits at the tenant, group and user level.
  • Replit Missions (17 November). Replit's launch post says "Your Missions agreement sets the spending limit and which actions require your permission", with the example of capping a mission at $1,000.
  • Claude pooled budgets (28 October). Anthropic gave Enterprise groups "one shared monthly budget", where a member "stops at whichever runs out first".
  • ChatGPT Wallet (13 November). OpenAI's Wallet runs in dots. OpenAI says the user reviews the items, total and card before each purchase, and that user-set spending limits may come later.

Each of these is a good step. Each also caps one maker's agents, for that maker's users. That's the gap this guide comes back to: a business doesn't run one maker's agents. It runs Claude Code for code, Codex for review, Copilot for email, a sales agent, a support bot. Every one of them has its own meter.

How can an AI agent spend your money?

An AI agent spends money in five main ways: model time (the thinking), compute (its computer and tools), paid actions in your apps, purchases it makes with a card, and subscriptions it signs up for. The first two show up on the AI tool's bill. The last three can show up anywhere.

1. Model time

Every time an agent reads, plans or writes, it uses a model. Bigger models and longer tasks cost more. This is what most "credits" pay for. Microsoft says a Cowork task's price depends on model use, context retrieval, tool calls and runtime.

2. Compute and tools

Agents that browse, run code or use a cloud computer pay for that time. Searches, tool calls and connectors often have a price too. In Vestra, the handbook says the rate is the same for model time, actions in your apps, the team's computer, searches and calls: one credit is one cent.

3. Paid actions in your apps

An agent working in your tools can trigger costs outside the AI bill. Sending 10,000 text messages costs money at your texting provider. Launching an ad set costs money at the ad platform. Running a big data export can cost money at your cloud provider. None of this shows up on the AI tool's meter.

4. Purchases with a card

Some agents can pay for things: a list of contacts, a software plan, a booth at a trade show. This is the riskiest kind of spending, because it's real money leaving your account, often to someone new.

5. Subscriptions and sign-ups

An agent that signs up for a free trial can leave you with a monthly bill when the trial ends. This is easy to miss, because the first charge comes weeks later.

The hidden multiplier: loops and retries

The cost that surprises people most is the agent that keeps trying. A task that fails, retries, fails again and retries again can use many times its normal cost. Without a per-task cap, nothing stops it until a daily or monthly limit kicks in, and by then the money is spent.

What kinds of spending caps are there?

There are eight common kinds of AI agent spending cap: per task, per day, per month, per user, per group, per agent, per workspace and per purchase. Each catches a different problem. A monthly cap stops a slow leak. A per-task cap stops one runaway job. A per-purchase cap stops one bad payment. Good setups use several together.

  • Per task. A limit on what one job may cost. It stops the runaway loop before it eats your month. Few AI tools offer it. Replit's mission budget is one of the closest, because it caps a whole goal.
  • Per day. A limit on what all agents may spend in a day. It catches a bad day early, before a monthly cap would.
  • Per month. The most common cap. It protects your total bill, but a single bad task can still burn a big share of it in an hour.
  • Per user. A limit on what one person's use may cost. Microsoft, Claude, ChatGPT, Lindy, Cursor and GitHub all offer some version.
  • Per group or team. A shared limit for a department or project. Claude's pooled budgets and Devin's Budgets work this way.
  • Per agent. A limit on one agent, whoever uses it. Tasklet and Devin Bots offer this. Perplexity says usage limits apply to the teammate's owner, not to a teammate.
  • Per workspace or organisation. The ceiling for everything. Almost every tool has one.
  • Per purchase or per card. A limit on one payment, or on a card the agent holds. Stripe Issuing, Visa and Mastercard all describe controls like this on the payment side.

On top of what is capped, there's how the cap behaves:

  • Hard stop or soft alert. Does the agent stop, or does someone just get an email?
  • Real time or delayed. Is the cap checked before each step, or a while later? Microsoft says Cowork limits "are evaluated asynchronously".
  • Stops new work or running work. Some caps block new tasks but let running ones finish. Lindy says "A task already running can finish past the cap."
  • What comes after. Does the agent wait for you, fall back to pay-as-you-go, or just stop with the work half done?

Keep these four questions in mind as you read the next sections. They tell you more about a tool's cap than its name does.

How does Microsoft Copilot cap agent costs?

Microsoft caps Copilot agent costs with spending policies in Microsoft 365. Policies cover the tenant, groups and users, each with a monthly limit in Copilot Credits. When a user hits the limit, they lose access to the agents until the next month. Microsoft says the checks run asynchronously, so a limit is not a real-time gate.

Microsoft has the most detailed cost controls of any maker in this guide, because Copilot is sold to big companies with finance teams. Here is how it works, from Microsoft's own pages.

Copilot Cowork and usage-based billing

Copilot Cowork is the part of Copilot that carries out multi-step tasks for you. It went generally available in June 2026. Microsoft's launch post lists "Spending limits at the tenant, group, and user levels" and says "Admins create scoped billing policies and define budgets, including user-level caps set inside group policies." It also promises "Customizable usage alerts" where admins set the thresholds and choose who gets notified, and "User-level pricing for each task, in credits."

The price is simple on paper: "PayGo is priced at $0.01 per Copilot Credit." Microsoft also sells credits in advance at a discount.

Microsoft Learn explains the details:

  • Levels. "The system supports user, group, and tenant policies." Specific users are scoped through security groups.
  • Monthly reset. Limits reset at the start of the month, 00:00 UTC, and unused credits don't carry over.
  • What happens at the limit. "When users hit the limit, they lose access to agents and services for the rest of the month until credits reset on the first of the month." They can request more credits.
  • Order of payment. A policy "always uses prepaid credits first" before pay-as-you-go.
  • Suggested size. Microsoft recommends setting "at least 2,000 credits per user per month", which is $20 at the pay-as-you-go rate.
  • Two choices. When admins set up a policy, the options are "Don't limit monthly spending" and "Limit monthly spending".
  • Alerts. "Alert emails begin when usage reaches the specified threshold and continue daily."

The catch: limits are not real time

Microsoft is open about this. Its Cowork access page says "Credit consumption and limit enforcement are evaluated asynchronously." Users "might be able to start one or more tasks after their limit is reached before enforcement takes effect." And, in Microsoft's own words: "Credit limits are a spending safeguard, not a real-time access gate."

There's also no way to see what a task will cost before you start it. Microsoft's page on checking cost answers the question "Can I check the cost before starting?" with "No." You can see the running total for a session afterwards.

Copilot Business and Autopilot

For Microsoft 365 Copilot Business, Microsoft's billing FAQ says "The default spending policy limits usage to 4,000 Copilot credits per user per month." When users reach it, they see a message and "can request more capacity from their administrator."

Copilot Autopilot, Microsoft's new always-on personal agent, needs a spending policy before it works at all. Microsoft Learn says to "set up a spending policy that covers Autopilot" and that "If a user isn't covered, or the policy runs out of credits, Autopilot doesn't work for that user, and Autopilot doesn't show the user an explanation." Admins can "Set a monthly limit, a per-user limit, or both, and add email alerts." If a user sits in more than one policy, the one with the highest limit applies. Our guide What is Microsoft Copilot Autopilot? covers the rest of Autopilot.

Copilot Studio agents

Agents built in Copilot Studio draw on prepaid capacity. Microsoft "allows some level of overage consumption, similar to a grace period", and "Enforcement is triggered when a tenant reaches 125% of their prepaid capacity." Then "A blocked agent stops responding, including in a conversation that's already in progress." With pay-as-you-go turned on, running out of prepaid credits doesn't block the agents; the extra is billed.

Agent 365 cost management

At Ignite on 17 November, Microsoft said Agent 365 lets admins "Apply spending limits and alerts at the user, group, organization, or resource level", "Track consumption across users, services, agents, and models" and "Control which models teams can use." Spend data can flow into chargeback reports. Microsoft's Book of News describes the spending policies, limits and alerts as a preview. The Agent 365 product page lists Agent 365 at $15 per user per month, billed yearly.

What this means for a small business

Microsoft's controls are strong if you live inside Microsoft 365 and have an admin to set them up. They cap Microsoft's agents, per person, per month, a little after the fact. They don't cap a single task, and they don't see the Claude Code or Codex work your developers run elsewhere. For the full comparison, see Vestra vs Microsoft Copilot.

How does Replit cap Missions and Agent spending?

Replit caps Missions with a budget in US dollars that you set before the mission starts, plus a list of actions the agent must ask about first and actions it must never take. When the budget runs out, the work stops, even if the goal isn't met. Outside Missions, usage limits cap spending beyond your monthly credits.

Replit Missions launched on 17 November 2026. It points a team of Replit agents at one goal, like "get 50 paying customers for this app", and lets them plan, build, test and keep going. That's exactly the kind of work that can run up a bill, so Replit made the budget part of the setup.

The mission agreement

Replit's launch post says: "Your Missions agreement sets the spending limit and which actions require your permission." Its examples: "you can cap the mission's spending at $1,000" and "require approval before the Agent contacts prospects." Replit adds that "These limits and approval requirements remain in place as the approach evolves", which means the agent can change its plan but not its budget.

Replit's docs describe the budget as "The maximum Replit AI usage for the mission," set "in USD." In the boundaries step, you "Specify which actions Replit must ask you about first and which actions it must never take."

What happens at the budget

Replit is honest about two things. First, "Work already running can take AI spending slightly over the budget before it stops." Second, "The budget does not guarantee that the mission will succeed within that amount." The launch post puts it plainly: "If the budget runs out first, work stops with the outcome still unresolved."

Account-wide limits

For the rest of Replit, usage limits do the capping. Replit's docs say "Usage limits cap your spending beyond your monthly credits." When reached, "usage-based services are blocked until the next billing cycle or until you increase the limit." Budgets are set "in increments of $500", and "Enterprise admins can give individual members a different spend limit than the workspace default."

Replit Agent uses effort-based pricing. Replit's own post says simple changes are "typically costing less than $0.25" per checkpoint, while bigger tasks are bundled into one checkpoint that "may cost more than $0.25."

What this means for a small business

Replit's mission budget is the closest thing in this guide to a per-goal cap set before the work starts, and the boundaries step is a good model for any agent: say up front what it must ask about and what it must never do. But the budget covers Replit's AI usage on that mission. It doesn't cover your other agents, and it's built for software outcomes.

How does Claude cap spending?

Claude caps spending with limits that owners set. On Team and seat-based Enterprise plans, owners can set an organisation-wide spend limit and, where available, a monthly limit for each member. Enterprise adds group limits and pooled group budgets. When a member hits their limit, their usage credits pause. Anthropic says you can slightly exceed a limit.

Anthropic has added cost controls in steps through 2026. In July it added spend-threshold alerts "at 75% and 90% of an org-level spend limit" and let admins "set model defaults and spend limits by group." On 28 October it added pooled group budgets, which "give a group one shared monthly budget". In Anthropic's words, "Every request counts against both, and a member stops at whichever runs out first." Admins get alerts at 50%, 75%, 95% and 100%.

Team and Enterprise spend limits

Anthropic's help centre says "Owners and Primary Owners of Team and seat-based Enterprise plans can set organization-wide spend limits." They "can also set individual monthly spend limits, where available, for each member", though "Spend limits for individual members aren't available to every Team organization."

At the limit: "Once a user reaches their defined spend limit, this will automatically pause their usage credits." When the organisation limit is reached, "you won't be able to use Claude, Cowork, or Claude Code again until the next billing period." And, like Replit and Microsoft, Anthropic warns: "It's possible to slightly exceed your defined spend limit."

On Enterprise, groups can have their own limits, and a member's "individual limit takes precedence over any group limit." There's also a Spend Limits API that can "set a spend limit on each Claude Enterprise member", for Enterprise organisations only.

Usage limits and usage credits

Paid Claude plans have usage limits that reset: a five-hour window and weekly limits "across all models." When you hit them, "With usage credits enabled, your work continues on usage credits instead of stopping." On individual plans, you can "Set a maximum amount you're willing to spend on usage credits each month", or choose "Set to unlimited." Auto-reload buys more credits when the balance drops below a threshold.

Claude Code

Anthropic's Claude Code docs say Team and Enterprise admins can "set spend limits at the organization, group, or individual member level", and Console users can "set workspace spend limits on total Claude Code spend." Inside Claude Code, the usage command shows a session's total cost, which Anthropic calls "an estimate." Members without billing access can request more credits, which "sends a request to your organization's admins." See Vestra vs Claude Code for how Claude Code fits in a team.

Claude Tag in Slack

Claude Tag, Claude's Slack app, bills differently. Anthropic says work in channels "bills to your organization's usage balance, not to individual seats," and "The spend limit is a cap you set on how much of that balance Claude Tag can use each month." Claude posts warnings in Slack at 75% and 90%. At the limit, "Claude stops and tells the requester in the thread that it couldn't finish." On invoiced Enterprise, Anthropic notes usage bills "with no upper bound until you set a spend limit."

The Claude API

On the API, Anthropic sets monthly caps by tier ("Start $500… Build $1,000… Scale $200,000"), and "Custom tier have no monthly spend cap." At the tier cap, "API usage pauses until 00:00 UTC on the first day of the next month." You can set your own lower limit, and "you can set custom spend and rate limits for each workspace."

Claude Duet and purchases

Claude Duet, Anthropic's background agent in research preview, handles purchases with an approval each time. Anthropic's safety page says Duet "waits for you to approve the action" and you "approve the purchase in your Link app", using a one-time card "limited to that merchant and amount." Paying or sending money can't be put under a standing "don't ask me again" rule. We found no general spending cap for Duet on Anthropic's pages.

What this means for a small business

Claude's controls are thorough for a Claude-only shop: organisation, group, member, workspace and Slack limits, plus a strong per-purchase approval in Duet. Like Microsoft's, they cap spend per period, not per task, and they only see Claude.

How does ChatGPT cap spending?

ChatGPT caps spending through workspace credits. On Business, when the credit pool is empty, features stop until an owner adds more. On Enterprise, admins can set a workspace overage limit, which starts at "No limit", and per-user credit limits with group and individual overrides. Codex and ChatGPT Work share the same credits and limits.

OpenAI's help centre pages on spend controls didn't load for us on 22 November, so these facts come from the pages that did: OpenAI's flexible pricing article, the ChatGPT learning site and OpenAI's announcements.

Business and Enterprise credits

For Business, OpenAI says: "If no credits are available in the workspace pool, the feature is blocked and users can ask a workspace owner to add more." Owners "can use automatic recharge to make sure the workspace has sufficient credits." We didn't find a cap on automatic recharge.

For Enterprise and Edu, "the workspace overage limit starts at No limit." Setting "A limit of 0 blocks new requests that use credits after the pool is empty." Admins "can use RBAC to set spend controls by group."

Per-user limits for Work and Codex

OpenAI's admin FAQ for ChatGPT Work says admins can "Establish a default per-user credit limit for the workspace", and that "ChatGPT Work and Codex share pricing, credits, and usage limits." An individual override "takes precedence over group and workspace defaults." If your agreement doesn't permit overages, "eligible usage stops when no available credits remain." One line is worth remembering: "Alerts notify; they don't stop spending."

Codex has limits per "five-hour period," and "Weekly limits may also apply." Plus and Pro users who reach their limit "can purchase additional credits." See Vestra vs Codex and Vestra vs ChatGPT Work.

Purchases: ChatGPT agent, dots and Wallet

OpenAI says ChatGPT agent is "trained to explicitly ask for your permission before taking actions with real-world consequences, like making a purchase." Dots, OpenAI's always-on agents, "start with built-in rules for when to act independently and when to ask for approval." ChatGPT Wallet, announced on 13 November and running in dots, works like this: "Before each purchase, the user reviews the items, total, and selected card," and it "requires a security check for every purchase." OpenAI describes user-set spending limits for Wallet as something that may come in future. We found no dots spending limit on OpenAI's pages. See Vestra vs OpenAI Dot.

The OpenAI API

Developers can "Set monthly API spend limits for your organization and projects," with an "Enforce a hard limit" switch. A hard limit returns an error once reached, and "Recorded spend can slightly exceed the configured amount." A spend alert "Sends a notification; API traffic continues."

What this means for a small business

ChatGPT's controls are per workspace and per user, with a sensible default that blocks work on Business when credits run out. Enterprise needs an admin to set a limit, because the default is none. Purchases get a review each time. As with the others, nothing caps a single task, and nothing sees your other agents.

How do other AI agents cap spending?

Other AI agents cap spending in very different ways. Tasklet and Devin can cap individual agents. Lindy caps each teammate's monthly credits. Perplexity caps a Max account's monthly charges. GitHub Copilot and Cursor cap users and organisations. Google caps Gemini Enterprise through a billing budget. For Manus, we found no spending cap on its own pages.

Tasklet

Tasklet, which gives a company AI agents it builds and manages one by one, has some of the most specific caps. Its docs say "Admins can set daily and monthly limits for every agent, or for specific ones." Threads have a default cap of "10,000 credits/day by default." When a limit is reached, "work pauses until it resets," and "New requests and automations pause until credits are available." Email alerts go out "at 80%, 95%, and 100% of your credit allowance."

Card purchases are another matter. Tasklet's own guide says "The Agent is instructed to obtain approval before filling a saved card," then adds: "This relies on the Agent following instructions; there is no enforced per-purchase approval gate or spending limit." It also says "Usage notes guide the Agent; they do not enforce spending limits or merchant restrictions," and "Anyone who can use the Agent can use its saved cards." That candour is useful. It's the clearest statement in this guide of the difference between asking an agent to behave and making it. Our page Vestra vs Tasklet compares the two in full.

Lindy

Lindy's docs say "Admins can set a monthly credit cap for each teammate," and "With no cap, a teammate can use the whole workspace balance." At the limit, "Lindy pauses new credit-using actions for that person," but "A task already running can finish past the cap." Lindy also "can pause a task that's using more credits than usual." Its pricing page lists top-ups at "$10 per 1,000 credits." We didn't find an approval step for purchases. See Vestra vs Lindy.

Manus

Manus bills in credits that refresh monthly, with daily and event credits on top. We found no spending cap, no page on what happens when credits run out, and no confirmation step for purchases on Manus's own help pages. Its cloud browser asks you to take over for SMS codes, CAPTCHAs and two-factor sign-ins. See Vestra vs Manus.

Devin

Cognition's Devin bills in ACUs (agent compute units). Its docs say "An organization-level limit caps the organization's monthly ACU consumption." At the limit, "New work billed to that organization is blocked until an administrator raises or removes the limit." Devin is moving these into Budgets, with "one limit for the account, shared pools for teams or projects, and per-user limits inside each pool." Devin Bots, the persistent manager bots launched on 17 November, can be capped too: "You can cap the Bot with an ACU limit or a spending limit in dollars," and "At the limit, the Bot stops until the limit resets or someone raises it."

Perplexity

Perplexity says "Every Max account has a monthly spending cap on total credit charges per billing period," $200 by default and adjustable "up to $5,000." At the cap, "No more charges are made until you update settings." Auto-refill is off by default. For Teammates, Perplexity says "Usage limits apply to the teammate's owner, not to a teammate or a Project." Its Comet browser assistant can be set to ask before actions, or to "Always Allow". See Vestra vs Perplexity Teammates.

GitHub Copilot

GitHub now bills Copilot by usage, with budgets per user, cost centre, organisation and enterprise. "When a user reaches an enforced budget limit, their access to Copilot features that consume AI credits is blocked." But for organisation, cost-centre and enterprise budgets, the "Stop usage when budget limit is reached" setting is "off by default", and without it "charges continue to accrue past the limit."

Cursor

Cursor asks you to set a spend limit the first time you use its cloud agents. When a member's limit is hit, "Paid additional usage stops for that specific user" while "Other team members continue unaffected." Enterprise can set "both team-level and member-level limits."

Google Gemini Enterprise

Google caps Gemini Enterprise usage above the included quota with "A Cloud Billing budget cap that stops paid usage." Google warns that "stopping usage can take some time to take effect," so charges can go over. Without a cap, "feature usage continues without restriction." There's a setting to limit any single user's quota, but Google says these limits don't cap pay-as-you-go usage after the pool runs out.

Personal agents: Grok Bot, Meta Muse, Hermes and OpenClaw

Personal agents mostly cap spending through the plan you pay for. You get a set amount of use for a monthly price, and the agent works for you alone. Open-source agents like OpenClaw and Hermes run on your own model keys, so the cap is whatever you set at your model provider. See Vestra vs Grok Bot, Vestra vs Meta Muse, Vestra vs Hermes Agent and Vestra vs OpenClaw.

AI agent spending limits compared, tool by tool

Most AI tools cap spending per month and per user, some cap per group or per agent, and almost none cap a single task. Purchases are usually handled by a review before each payment rather than a cap. Here is each tool in one place, from the makers' own pages, checked 22 November 2026.

Microsoft Copilot (Cowork, Autopilot, Studio, Agent 365)

  • What's capped: tenant, group and user, monthly, in Copilot Credits ($0.01 each).
  • Per task: no cap found. Per-task cost is shown in credits after the fact.
  • At the limit: access lost until the 1st of the month; users can request more.
  • Real time? No. Microsoft says limits are "evaluated asynchronously".
  • Alerts: yes, thresholds set by admins, daily emails.

Replit (Agent and Missions)

  • What's capped: each mission, in US dollars, set up front; account usage beyond monthly credits; per-member limits on Enterprise.
  • Per task: the mission budget caps a whole goal.
  • At the limit: work stops, even with the goal unmet. Running work can go slightly over.
  • Approvals: you choose actions the agent must ask about first and must never take.

Claude (Team, Enterprise, Claude Code, Claude Tag, API)

  • What's capped: organisation, group, pooled group and member (where available), monthly; Claude Tag monthly and per channel; API by tier and workspace.
  • Per task: no cap found.
  • At the limit: usage credits pause until the next period; Claude Tag stops and says so in the thread.
  • Alerts: 75% and 90% (org), 50% to 100% (pooled budgets).
  • Purchases: Duet asks every time, with a one-time card for that merchant and amount.

ChatGPT (Business, Enterprise, Work, Codex, dots)

  • What's capped: workspace pool; Enterprise overage limit (default "No limit"); per-user limits with group and individual overrides.
  • Per task: no cap found.
  • At the limit: Business features blocked until an owner adds credits.
  • Alerts: "Alerts notify; they don't stop spending."
  • Purchases: ChatGPT agent asks first; Wallet in dots reviews every purchase.

Tasklet

  • What's capped: each agent, daily and monthly; threads 10,000 credits a day by default.
  • At the limit: work pauses until the limit resets.
  • Purchases: relies on the agent following instructions, with no enforced per-purchase gate (Tasklet's words).

Lindy

  • What's capped: each teammate (person), monthly.
  • At the limit: new credit-using actions pause; running tasks can finish past the cap.

Devin

  • What's capped: organisation and user, monthly, moving to Budgets with team pools; each Bot by ACUs or dollars.
  • At the limit: new work blocked; a Bot stops until the limit resets or is raised.

Perplexity (Max, Teammates, Comet)

  • What's capped: the Max account's monthly charges ($200 default, up to $5,000).
  • Per teammate: no. Limits apply to the owner.
  • At the limit: no more charges until you change settings.

GitHub Copilot, Cursor, Gemini Enterprise

  • GitHub: budgets per user, cost centre, organisation and enterprise; the stop setting is off by default above user level.
  • Cursor: spend limits per member and team; paid usage stops for that member.
  • Gemini Enterprise: a Cloud Billing budget cap that stops paid usage, with a delay.

Manus

  • What's capped: monthly credits by plan. No spending cap found on Manus's pages.

Vestra

  • What's capped: any single task and the whole day (up to $100 a day), for every teammate on every engine, live today.
  • Per task: yes. The handbook's example sets $5 per task.
  • At the limit: the agent stops at a sensible point, says what it finished and what's left, and waits for you.
  • Money: spending over a limit always waits for your yes.
  • Receipts: every credit itemised, by teammate and by piece of work.

What do today's caps miss?

Today's caps miss three things: the single task, the moment, and the team. Most caps are monthly, so one bad task can burn the month before anyone notices. Several are checked late, so work runs past the line. And each cap only sees one maker's agents, so a business using several tools has several budgets and no total.

They cap the month, not the task

A monthly cap is a seatbelt for the whole year. It doesn't stop one job that loops for an hour. In this guide, only Replit's mission budget comes close to a per-job cap, and it covers a whole goal rather than a single step. Microsoft says you can't see a Cowork task's cost before you start it. Most tools show you a task's cost only after it's done.

They catch up later

Microsoft calls its limits "a spending safeguard, not a real-time access gate." Google says stopping "can take some time." Replit, Anthropic and OpenAI all say spend can slightly exceed a limit. Lindy lets a running task "finish past the cap." None of this is wrong. It's how metered systems work. But it means a cap is a fence with a gap at the bottom, and the gap is widest exactly when an agent is moving fast.

They count per person, not per team

This is the big one, and it's the same pattern this blog keeps coming back to. Personal agents like Claude Code, Codex, Grok Bot, Meta Muse, OpenAI Dots and Hermes are built for one person, and their caps are built the same way: per user, per seat, per account. Even the tools that now ship agent teams, like Devin Bots, Copilot Workspaces and Perplexity Teammates, cap their own maker's agents on their own bill.

Picture a ten-person company in 2026. The developers use Claude Code and Codex. The founder uses Copilot. Sales uses a Lindy agent. Support uses a bot in Slack. Every tool has a cap. Every cap is set somewhere different, in different units (Copilot Credits, ACUs, usage credits, plain dollars), with different alert levels and different rules for what happens at the line. Nobody can answer the simple question: "How much did our agents spend today, and on what?"

That person in the middle, reconciling five bills, is the same person who pastes context between five agents and chases five results. We wrote about it in How to manage multiple AI agents without becoming their manager. Spending is one more thing you end up carrying by hand, unless the agents work in one place under one set of rules.

Purchases sit somewhere else

AI credits and real purchases are usually two different systems. Your AI tool caps credits. Your card, or a payment rail, caps purchases. An agent that's well inside its credit budget can still spend $480 on a trade show booth if its card allows it. A good setup covers both.

How do payment cards and rails cap agent purchases?

Payment companies now offer controls built for agents. Stripe Issuing offers per-agent spend limits, merchant category controls and real-time approval of each purchase. Visa Intelligent Commerce lets a person set spend parameters inside an instruction. Mastercard Agent Pay defines agent permissions and spending limits up front. These cap the money itself, wherever the agent spends it.

  • Stripe Issuing for agents. Stripe says it offers "per-agent spend limits, merchant category controls, and custom authorization rules", set "per transaction or over a rolling period." Single-use cards "are automatically invalidated after use," and "every purchase attempt triggers a real-time authorization request before the charge goes through."
  • Stripe shared payment tokens. For agents buying from sellers, Stripe lets you "Set usage limits, including currency, maximum amount, and expiration window," and revoke the token.
  • Visa Intelligent Commerce. Visa describes "Consumer-defined spend parameters" within an instruction, which Visa checks across the whole instruction, and checks that the merchant matches.
  • Mastercard Agent Pay. Mastercard says "Agent permissions and spending limits are defined upfront to bound authority." Its page didn't say how limits are enforced.

Card-level controls are the strongest guard you can have for purchases, because the bank or payment network enforces them, not the agent. An agent can't talk a card into a higher limit. That's why it matters whether a tool's purchase approval is enforced by the system or only by instructions to the model, as Tasklet's guide spells out.

How do you set spending limits for your AI agents?

Set spending limits for your AI agents in layers: cap each task, cap each day, cap each agent's card, set a monthly ceiling, and require a person's yes for anything over a cap, anything new and any purchase. Then read the receipts weekly and adjust. The ten rules below work whatever tools you use.

1. Cap every task

Decide what a normal job should cost and set a cap a little above it. A research task that usually costs 30 cents doesn't need a $20 ceiling. If your tool has no per-task cap, keep big jobs small: split them into steps you can check.

2. Cap every day

A daily cap catches a bad day while it's still a day. Set it at two or three times a normal day's spend. Monthly caps are still useful, but they're the last line, not the first.

3. Give each agent its own card and limit

If agents buy things, give each its own card with its own monthly limit, never your personal card or the company's main card. Use single-use or merchant-locked cards where you can. Stripe, Visa and Mastercard all describe controls like this.

4. Ask before anything big

Have the agent check with you, with the cost, before any job that's much bigger than usual. It's cheaper to say "start with the 100 closest" than to undo a $40 job you didn't mean to run.

5. Make the yes real, not a promise

Prefer approvals the system enforces over instructions the model is asked to follow. "The agent is told to ask before paying" is not the same as "the payment can't go through without a yes."

6. Always ask for the first time

The first payment to a new merchant, the first email to a new person and the first sign-up for a new service should always wait for a person. New is where most mistakes and most scams start.

7. Turn alerts into stops

An alert at 80% is useful. An alert at 100% with no stop is a receipt for money already gone. Wherever your tool lets you, pick "stop" over "notify" at the limit. GitHub's organisation budgets, for example, don't stop usage unless you turn that setting on.

8. Watch for runaway loops

Ask for a cost per task, not just a total. If one job costs ten times the others, find out why. Lindy's ability to pause a task "using more credits than usual" is the right idea.

9. Read the receipts every week

Spend five minutes a week on what each agent spent and what it got done. The goal isn't to cut costs. It's to know what each dollar bought.

10. Put the rules in one place

If you use several tools, write your rules once and set them in every tool the same way. Better still, run your agents somewhere that applies one set of rules to all of them.

What should always wait for a person's yes?

Three kinds of action should always wait for a person's yes, however small and however trusted the agent: spending over a limit, deleting something, and contacting someone new. Many teams add a fourth: any payment at all to a new merchant. These are the actions that are hardest to undo.

  • Spending over a limit. Above a task's cap, the day's cap or a card's limit, the agent stops and asks. The question should come with the amount and the reason.
  • Deleting something. A file, a page, an email, a record. Gone is gone.
  • Contacting someone new. The first email, call or message to anyone you haven't dealt with. It speaks for your business, and it's how a mistake reaches the outside world.
  • Paying someone new. The first payment to a merchant the agent hasn't paid before.

The best tools make these questions quick to answer. A good approval gives you the question in one line, the facts that matter, a suggestion and the choices, so you can say yes from your phone. Our guide Delegate work to AI without losing control goes deeper on approvals and freedom levels.

How much should an AI agent be allowed to spend?

An AI agent should be allowed to spend about two to three times what its normal work costs: a per-task cap a little above a typical task, a daily cap a little above a typical day, and a card limit sized to the purchases its job needs. Start low, then adjust from the receipts.

Here is one way to size it, as a made-up example for a small business. These numbers are an illustration, not a quote from any tool.

  • A research agent that reads websites and writes summaries. Typical task: under 50 cents. Per-task cap: $2. Daily cap: $10. Card: none, because research doesn't need to buy anything.
  • A sales agent that follows up with leads by email. Typical task: under 20 cents. Per-task cap: $1. Daily cap: $10. Card: $100 a month for a contact list, with the first purchase from any new vendor waiting for a yes.
  • A coding agent that builds and tests features. Typical task: a few dollars. Per-task cap: $10. Daily cap: $40. Card: none.
  • A bookkeeping agent that sends invoices and chases late payments. Typical task: under 20 cents. Per-task cap: $1. Daily cap: $5. Card: none, and refunds always wait for a person.

Add the daily caps and you get a team ceiling of $65 a day in this example. That single number is the one most businesses can't see today, because each agent's cap lives in a different tool.

What happens when an agent hits its cap?

When an agent hits its cap, one of four things happens: it stops and loses access until the period resets, it pauses and waits for someone to raise the limit, it keeps going on pay-as-you-go, or it stops with the work half done. The best agents stop at a sensible point, say what's done and what's left, and wait.

Here is what each maker says happens at the line:

  • Microsoft Cowork: users "lose access to agents and services for the rest of the month until credits reset on the first of the month." Autopilot simply stops working for that user, without an explanation.
  • Copilot Studio: "A blocked agent stops responding, including in a conversation that's already in progress."
  • Replit Missions: "work stops with the outcome still unresolved."
  • Claude: usage credits pause; at the organisation limit you can't use Claude, Cowork or Claude Code until the next billing period. Claude Tag "stops and tells the requester in the thread that it couldn't finish."
  • ChatGPT Business: "the feature is blocked and users can ask a workspace owner to add more."
  • Tasklet: "work pauses until it resets."
  • Lindy: new actions pause, but a running task can finish past the cap.
  • Devin Bots: the Bot "stops until the limit resets or someone raises it."

The worst outcome isn't overspending by a few cents. It's an agent that stops halfway through something that matters, like a half-sent email campaign or a half-finished data migration, and nobody knows. Whatever tool you use, check what a stop looks like before you rely on it. Ask: will I be told, will I know what was done, and can it pick up where it left off?

How does Vestra cap what its AI team spends?

Vestra caps every agent with the same rules, whichever engine it runs on: a cap on any single task, a cap on the whole day of up to $100, a check before unusually big tasks, your yes before spending over a limit, and a receipt for every credit. The caps, the yes and the receipts are live today.

Vestra is one office where a team and its AI agents work together. You get a chief, an agent that runs your other agents and answers to you, and teammates hired for whole jobs. Each teammate runs on the engine that suits its job: Vestra's own, Claude Code or Codex today. That's the core difference from every tool above. Claude Code and Codex are excellent, but each is built for one person and caps its own spending on its own bill. In Vestra, a builder on Claude Code and a reviewer on Codex work side by side as teammates, under one chief, one set of office rules and one receipt.

The same caps on every engine

The handbook is plain about it: "No engine can raise its own limit, approve its own payment or give itself access to an app." The office rules are "the same whichever engine a teammate runs on": the same yeses, the same limits, sealed logins and the same receipts. A teammate on Claude Code and one on Codex hit the same per-task cap and the same daily cap.

Per-task and daily caps

You set "a cap on any single task and a cap on the whole day, up to $100 a day." The handbook's example sets a single task at $5. When a cap is reached, "work doesn't fail halfway. The agent stops at a sensible point, says what it finished and what's left, and waits. Raise the cap for the day, or let it pick up tomorrow."

A check before big tasks

The handbook adds one more guard: "Before starting anything unusually large, by default anything over 50 credits, your chief tells you what it will cost and asks." Its example: a researcher wants to read the websites of 400 clinics, about 240 credits ($2.40), when most of its tasks cost under 30. Your chief offers three choices: go ahead, start with the 100 closest, or not now. That's the per-task estimate Microsoft says Cowork can't give you before a task starts.

Your yes on money

Spending over your limit is one of three things that always wait for a person in Vestra, along with deleting something and contacting someone new. Each question arrives the same way: one line, the facts, your chief's suggestion as choice number 1, and the options. You can reply "1" from your phone.

Every credit on a receipt

In Vestra, one credit is one cent, "the same for everything an agent uses: model time (its thinking), actions in your apps, its computer, searches and calls." Every piece of work comes with a receipt showing what it cost and which teammate spent it. The handbook's example: a renewal invoice that cost 19 credits (14 for model time, 2 for actions, 3 for four minutes of computer), or $0.19. Ask your chief "what did Scout cost this week, and what did we get for it?" and you get an answer you can check line by line.

What's live and what's coming

Vestra's handbook sets out plainly what works today and what's being built. Live today for one person and their team: a chief and teammates, a cloud computer you can watch and take over, calls and email from your chief's own number and address (calls from Pro up), 2,000+ connected apps, a memory that learns, any engine per teammate (Vestra's own, Claude Code or Codex), and "Daily and per-task caps, your yes on money, every credit itemised."

Coming later: a vault for every agent, with a card and sealed logins and a monthly limit each person sets. In the handbook's picture, a growth teammate's card has a $300 monthly limit; it pays for what the job needs inside the limit, and anything over the limit or anything new waits for a yes. That per-agent card is on the roadmap, not live today. So are a chief for every person on a team, chiefs working together, spaces and the walkable office.

Plans and credits

Plans start at $19 a month (Lite, up to 3 teammates, 2,000 credits a month), then Pro at $97 (20 teammates, 10,000 credits), Max at $197 (50 teammates, 20,000 credits) and Ultra at $497 (150 teammates, 50,000 credits), with lower prices billed yearly. Unused credits roll over, you can top up any time, and every new office starts with 1,000 credits free. In India, plans are billed in rupees with GST included. Because one credit is one cent on every engine, a plan's credits are also its budget: 2,000 credits is $20 of work.

A week of capped spending in Vestra

Here is a made-up week for a one-person business, to show how the caps work together. Ines runs a small online shop selling hand-dyed yarn from Porto. She uses only what's live today: her chief, four teammates, per-task and daily caps, her yes on money and receipts. She set a $3 cap per task and a $15 cap per day.

Monday

Ines tells her chief, which she named Linho: "I need someone to answer customer emails, someone to watch what other yarn shops charge, and someone to keep my product pages fresh." Linho suggests four teammates and their engines, and she says yes. Her support teammate and her researcher run on Vestra's own engine, her writer on Claude Code, and a reviewer that checks the writer's pages on Codex. Monday's receipt: 212 credits, $2.12.

Tuesday

The researcher wants to read the price pages of 300 yarn shops across Europe. That's well over 50 credits, so Linho checks first: about 180 credits, $1.80, when most research tasks cost under 30. Ines picks "start with the 50 closest rivals". The task costs 34 credits.

Wednesday

A product-page task gets stuck. The writer keeps reworking a page that the reviewer keeps sending back, because two colour names clash with the shop's style guide. At $3 the task stops. Linho tells Ines what's done (four pages finished and checked) and what's left (one page, the clash). She fixes the style guide in one line. The lesson sticks, and the page finishes the next morning. Without a per-task cap, that loop could have run until the monthly credits ran out.

Thursday

A big wholesale enquiry comes in from a shop she has never dealt with. The support teammate drafts a reply for Linho to send from its own email address, but writing to someone new counts as contacting someone new, so it waits for Ines. She reads it on her phone in the queue at the post office and replies "1".

Friday

Ines asks Linho: "What did the team cost this week, and what did we get?" The answer comes back as receipts by teammate: the researcher, the writer, the reviewer and the support teammate, each with its credits and its work. The week came to $11.40, and no single day came near her $15 cap. She can see that the reviewer on Codex cost less than the writer on Claude Code, and exactly what Wednesday's stuck page cost before the cap stopped it.

Notice what Ines never did: she never set a budget in four places, never reconciled four bills in four units, and never wondered which agent spent what. That's the point of one office.

Common mistakes with AI agent budgets

The most common mistakes with AI agent budgets are relying on alerts instead of stops, setting only a monthly cap, leaving the default at "no limit", sharing one card across agents, and having a different budget in every tool. Each one leaves a gap that a busy or looping agent will find.

  • Trusting alerts. "Alerts notify; they don't stop spending," as OpenAI puts it. Check whether each limit stops or only warns.
  • Only a monthly cap. It protects the year, not the afternoon. Add a daily cap and, if you can, a per-task cap.
  • Leaving the default. ChatGPT Enterprise's overage limit starts at "No limit". GitHub's organisation budgets don't stop usage unless you turn that on. Claude Tag on invoiced Enterprise has "no upper bound until you set a spend limit."
  • One card for every agent. If five agents share a card, you can't tell who spent what, and one bad agent can use the whole limit. Tasklet's guide notes "Anyone who can use the Agent can use its saved cards."
  • Instructions instead of controls. Telling an agent "always ask before buying" is not a spending limit. A cap the system enforces is.
  • A budget per tool and no total. Five tools means five caps in five units, and no one number for the team.
  • No owner. Someone has to read the receipts. If nobody does, the caps are guesses.
  • Caps so tight nothing gets done. A cap that stops every task trains people to raise it to "unlimited". Size caps from real receipts, then adjust.

Frequently asked questions

What is an AI agent spending limit?

It's a cap on how much an AI agent can spend before it stops or asks a person. It can apply per task, per day, per month, per user, per group, per agent, per workspace or per purchase. A real limit stops the spending; an alert only warns you.

How do I stop AI agents from spending my money?

Use layers: cap each task, cap each day, give each agent its own card with a limit, and require a person's yes for spending over a cap, anything new and any first payment. Then read the receipts every week. Prefer controls the system enforces over instructions to the agent.

Can I set a spending limit in Microsoft Copilot?

Yes. Admins set spending policies in Microsoft 365 for the tenant, groups and users, with monthly limits in Copilot Credits at $0.01 each and email alerts. Microsoft says limits are checked asynchronously, so users may start tasks after the limit before it takes effect. Copilot Business's default policy is 4,000 credits per user per month.

Does Copilot Autopilot have a spending limit?

Autopilot needs a spending policy that covers it before it works. Admins can set a monthly limit, a per-user limit or both, plus email alerts. Microsoft says if a user isn't covered or the policy runs out, Autopilot doesn't work for that user and doesn't show an explanation.

How does Replit Missions cap spending?

You set a budget in US dollars for each mission before it starts, and choose which actions Replit must ask you about first and which it must never take. When the budget runs out, work stops even if the goal isn't met. Replit says running work can go slightly over before it stops.

Can Claude admins set spending limits?

Yes. On Team and seat-based Enterprise plans, owners can set organisation-wide spend limits and, where available, monthly limits for each member. Enterprise adds group limits and pooled group budgets. Anthropic says it's possible to slightly exceed a limit, and alerts go out as spend approaches it.

Can I cap Claude Code costs?

Yes. Team and Enterprise admins can set spend limits for Claude Code at the organisation, group or member level, and Console users can set workspace spend limits on total Claude Code spend. Claude Code shows a session's estimated cost with its usage command.

Can I limit ChatGPT or Codex spending?

Yes, on business plans. Business workspaces stop when the credit pool is empty, until an owner adds more. Enterprise admins can set a workspace overage limit, which starts at "No limit", plus per-user credit limits with group and individual overrides. ChatGPT Work and Codex share the same credits and limits.

Do AI agents ask before buying things?

Many say they do. OpenAI says ChatGPT agent asks before purchases, and Wallet in dots reviews every purchase. Claude Duet waits for your approval and uses a one-time card. But Tasklet's own guide says its purchase approval relies on the agent following instructions, with no enforced gate. Check whether a tool's approval is enforced or only instructed.

What's the difference between a spending limit and a usage limit?

A usage limit controls how much you can use what you already pay for, like Claude's five-hour window or Codex's weekly limit. A spending limit controls how much extra you can be charged beyond that. Some tools switch to paid credits when usage runs out, so you need both.

Why do agents sometimes go over their spending limit?

Because many limits are checked a little after the spending happens, and running work is allowed to finish. Microsoft, Replit, Anthropic, OpenAI and Google all say spend can slightly exceed a limit or that stopping takes time. A per-task cap and a check before big tasks narrow the gap.

What is a per-task spending cap?

It's a limit on what one piece of work may cost. It stops a single runaway job, like an agent stuck retrying, before it uses a big share of your monthly budget. Few AI tools offer one. Vestra lets you set a cap on any single task, live today.

How much should I let an AI agent spend per day?

About two to three times what a normal day of its work costs. Start low, read the receipts for two weeks, then raise the cap where good work is being blocked. In Vestra, you can set a daily cap for the whole team of up to $100 a day.

How do I see what each AI agent spent?

Most tools show spend per user or per workspace, and some show a task's cost after it's done. If you use several tools, you'll have to add their bills together. In Vestra, every piece of work comes with a receipt that shows what it cost and which teammate spent it, on every engine.

Can I give an AI agent its own credit card?

Payment companies now support it. Stripe Issuing offers per-agent spend limits, merchant controls and single-use cards, and Visa and Mastercard describe agent spending limits too. In Vestra, a card in each agent's vault with a limit each person sets is on the roadmap, not live today.

How does Vestra cap spending across Claude Code and Codex?

Every teammate in Vestra works under the same office rules whichever engine it runs on: the same per-task cap, the same daily cap, the same yes before spending over a limit, and the same receipts. One credit is one cent on every engine, so a teammate on Claude Code and one on Codex show up on one receipt in one unit.

Glossary

  • Agent. An AI system that does work on its own, step by step, using tools.
  • ACU. Agent compute unit, the unit Cognition's Devin bills in.
  • Alert. A message sent when spend reaches a set level. It doesn't stop anything.
  • Approval. A person's yes before a specific action goes ahead.
  • Asynchronous enforcement. A limit that is checked a little after the spending happens, so work can run past it for a while.
  • Budget. A planned amount for a period. Some tools enforce it as a hard limit; others only track against it.
  • Chief. In Vestra, the agent that runs your other agents, speaks for you and answers to you.
  • Copilot Credit. Microsoft's billing unit for usage-based Copilot features, $0.01 each pay-as-you-go.
  • Credit. A billing unit for AI work. In Vestra, one credit is one cent.
  • Daily cap. A limit on how much can be spent in one day.
  • Engine. What an agent runs on, such as Vestra's own, Claude Code or Codex.
  • Hard limit. A cap the system enforces, stopping work at or near the line.
  • Mission. In Replit, a goal that a team of Replit agents works toward, with a budget and boundaries set up front.
  • Overage. Spending beyond the credits or usage included in a plan.
  • Pay-as-you-go. Billing for use as it happens, after included credits run out.
  • Per-task cap. A limit on what one piece of work may cost.
  • Pooled budget. One shared budget for a group, as in Claude Enterprise.
  • Receipt. A record of what a piece of work cost, who did it and who said yes.
  • Single-use card. A card number that works once and then stops, used to limit agent purchases.
  • Spend limit. A cap on total spend for a period.
  • Teammate. In Vestra, an agent hired for a whole job, on the engine you pick.
  • Usage limit. A limit on how much of an included plan you can use in a window, like five hours or a week.

Keep reading

Sources

All checked on 22 November 2026.