Claude Code · CodexSpend limits for AI coding agents

Put your agents
on a leash.

Leash sits between your agents and the model provider. It checks every call against your budget, catches runaway loops and gives you a real off switch, before the bill arrives.

Start your 3-day trial See how it works

Up to 71%smaller JSON in our fixture benchmark

0 promptsstored. Usage metadata only

1 commandto protect an agent

Illustrative run

~/checkout · leash run --cap 5 -- claude

Fix the flaky checkout test. Keep going until it passes.

⏺ Bash(pnpm test checkout)

✗ 1 failing · expected 200, received 500

⏺ Edit(src/cart/total.ts)

⏺ Bash(pnpm test checkout)

✗ 1 failing · expected 200, received 500

… same edit, same test, attempt 8

Session spend$2.31 / $5.00 cap

LeashPaused

Repeat loop detected

Same tool call 8× in 3 minutes. Session paused at $2.31.

›Ask your agent anything protected

Loop caught at 8 repeatsAlert sent to Slack

Where agent budgets quietly disappear.

Re-running the same failing test 40 timesRe-reading the whole repo every turnRetrying a 429 in a tight loopSub-agents spawning sub-agentsPasting 20,000 lines of logs into contextPolling a build that never finishesFix, revert, fix againRe-running the same failing test 40 timesRe-reading the whole repo every turnRetrying a 429 in a tight loopSub-agents spawning sub-agentsPasting 20,000 lines of logs into contextPolling a build that never finishesFix, revert, fix again

Running all night with nobody watchingA full JSON dump to read one fieldThe biggest model for a one-line typoA CI job stuck retrying foreverRewriting the same file in circlesSix agents doing one agent’s jobContext windows packed to the limitRunning all night with nobody watchingA full JSON dump to read one fieldThe biggest model for a one-line typoA CI job stuck retrying foreverRewriting the same file in circlesSix agents doing one agent’s jobContext windows packed to the limit

Leash

your machine

LOCAL-FIRST · YOUR KEYS · YOUR LIMITS

Local keys stay on your machine. You keep control.

Leash runs as a local proxy next to your agent. Budgets are enforced on your machine, from a local ledger, so your limits hold even when the dashboard is out of reach.

Keys stay on your machine

In local mode provider keys stay in your OS keychain or encrypted local vault, and are sent only to the selected provider.

Metadata, not prompts

Only usage metadata syncs to the dashboard. Prompts and responses are not logged; optional code memory stays on your machine.

Limits that work offline

Caps and the kill switch live on your device. A network or dashboard outage doesn’t switch them off.

How it works

How Leash stops a runaway agent.

1
Your agent makes a call
Claude Code, Codex or your own SDK app sends its request through Leash instead of straight to the provider.
2
Leash reserves the cost first
Before anything is forwarded, a conservative estimate is reserved against your session, daily and repo budgets.
3
It spots the loop
Your configured loop and spend-rate rules stop repeated calls and runaway bursts.
4
You decide what’s next
Get an alert in Slack, Discord or email. Resume, add budget, or keep it stopped.

Leash paused your sessionRepeat loop · $2.31 of $5.00

~/checkout · leash run --cap 5 -- claude

Fix the flaky checkout test. Keep going until it passes.

⏺ Bash(pnpm test checkout)

✗ 1 failing · expected 200, received 500

⏺ Edit(src/cart/total.ts)

⏺ Bash(pnpm test checkout)

✗ 1 failing · expected 200, received 500

… same edit, same test, attempt 8

Session spend$2.31 / $5.00 cap

LeashPaused

Repeat loop detected

Same tool call 8× in 3 minutes. Session paused at $2.31.

›Ask your agent anything protected

Loop caught at 8 repeatsAlert sent to Slack

Anatomy of a runaway session

Three minutes from bug to bill. Leash is in the middle.

00:00

Session starts
with a $5 cap

00:42

Agent hits a
failing test

03:10

Same fix,
eighth attempt

03:11

Leash pauses
the session

Then

You resume
or kill it

Without a limit, a loop like this keeps going until someone notices, often when the invoice does. Leash stops it at the rule you set and tells you exactly why. Defaults: $20 per session, $100 per day, a loop after 8 repeats in 3 minutes.

Setup

Four steps. About five minutes.

Use your existing agent and provider authorization. Leash adds protection around supported API traffic.

1
About a minute
Install the CLI
Download the checksum-verified Leash CLI and install it with npm. Node 22.13 or later is all you need. No Docker, no database.
Run one install command.
2
One click
Sign in
Approve the device code in your browser. Your workspace starts with a $100 daily budget you can change any time.
Approve the code in your browser.
3
Your rules
Set your caps
Choose session, daily and repo limits. Every budget blocks by default. Switch any rule to alert-only if you prefer.
Pick the numbers you’re comfortable with.
4
Protected
Run your agent
Same agent, same provider key, now with a limit. Watch every session’s spend live in the dashboard.
leash run --cap 5 -- claude

Built into Leash

Your agents, with the controls you need.

Explore the controls, reports and local tools in one place.

Spend caps, kill switch and loop detection

Hard spend caps
Reserve cost before dispatch with session, daily and repository limits.
Kill switch and loop detection
Stop a session or every agent, with explicit resume and crash recovery.

Tool compression, model routing and local memory

Automatic tool-output compression
Compact supported tool results reversibly while keeping your instructions.
Approved model routing
Choose a cheaper compatible model under your allowlist, with manual pins.
Local code memory and MCP
Index incrementally and retrieve relevant code without uploading your repository.

Token X-ray, spend insights and Burn Reports

Token X-ray
Rank tool patterns and inspect optimization metadata.
Spend insights and Burn Reports
Break down models, providers and cache use; export or share report totals.
Invoice reconciliation
Compare your usage export with a provider CSV on Pro.

Conversation handoff and guided setup

Conversation monitoring and handoff
Review context pressure, download a handoff and copy the next conversation prompt.
Guided agent connection
Save your setup, approve the device and install with backup and restore.

Devices, shared budgets, virtual keys and alerts

Devices, members and remote controls
Use scoped device access, invitations, revocation and queued controls.
Atomic shared budgets and virtual keys
Opt into one Postgres admission authority across participating devices.
Email, Slack and Discord alerts
Choose supported plan channels, safety notifications and opt-in weekly reports.

Network controls, memory scaling and encrypted keys

Governed network
Enforce destination allowlists inside the managed Docker runtime.
Bounded memory auto-scaling
Grow supported container memory only up to your configured maximum.
Local encrypted credentials
Keep provider keys in your OS keychain or private encrypted vault.

Let your agent handle the setup.

Connect once. Keep your caps, controls and cost insights together. Give your agent the setup instructions, or take the terminal route yourself.

Build my personal setup

Your terminal · after installing the Leash CLI

leash connect --agent auto --api https://useleash.dev --mode device

# Then start a protected conversation
leash run -- YOUR_DETECTED_AGENT

# Stop every local session
leash kill --all

In 25 seconds

See it from your side.

Who it’s for

Is Leash right for you?

A great fit if you

Run Claude Code or Codex on API keys
Let agents work overnight, in CI or unattended
Build on OpenAI, Anthropic or supported OpenRouter Chat models
Have had a surprise API bill, or want to avoid one

Less of a fit if you

Only use flat-rate ChatGPT or Claude subscriptions
Need image, audio or web-search fees metered. Leash covers text and tool calls
Can’t route agent traffic through a local proxy or gateway

Pricing

A small price for a hard stop.

Starter and Pro include a 3-day trial. Card required. The trial has 1 user, 1 device, spend caps, a kill switch and 3 days of history. No team, reports or alerts during the trial.

Starter

$11/month

A hard boundary for your own agents.

1 user · up to 3 devices
Spend caps, kill switch and loop detection
Burn Reports, Token X-ray and handoff
3 scoped virtual keys
30 days of history
Email alerts

Try Starter for 3 days

Pro

Most teams

$22/month

One control room for a small team.

Up to 5 members · 10 devices
Everything in Starter
Slack and Discord alerts + Slack controls
Shared team budgets
10 scoped virtual keys + audit export
1 year of history
Invoice reconciliation report

Try Pro for 3 days

Enterprise

Let’s talk

A custom agreement for your organization.

Unlimited seats and devices by agreement
SSO / SAML by arrangement
Dedicated support
Custom terms

Local tools included with Starter and Pro

Automatic tool-output compression, approved model routing, local code memory with MCP, and encrypted provider credentials. Governed network and bounded memory scaling are available in the supported local Docker runtime.

Prices in USD. Leash costs are separate from your model provider bill. Compare plans in detail

FAQ

The questions we hear most.

Want the details? The documentation covers every limit, command and data boundary.

What does Leash actually protect?

Requests that go through the Leash proxy or gateway. Leash reserves a conservative cost estimate before a call reaches the provider, then meters the response. You can set session, repository and daily budgets, stop one session, or pause the whole workspace. Requests that bypass Leash are outside its protection.

Do I need to change my agent or share my prompts?

No. Keep Claude Code, Codex or your OpenAI- or Anthropic-compatible app. Leash ships installers and a command wrapper. Your dashboard receives usage metadata (tokens, models, cost), not prompts, responses or source code. Optional code memory stays local.

Is the number in Leash my actual invoice?

It’s an API-equivalent estimate based on reported usage and a published price table. Your provider’s invoice is the source of truth. Pro includes a report that reconciles your usage with a provider CSV export.

Can a team share one budget across devices?

Yes, when you turn on shared admission. Participating devices reserve against the same budget before forwarding calls. In ordinary local mode, each device enforces its own limits.

Does it work with ChatGPT or Claude subscriptions?

Leash meters API traffic. Use Claude Code with an Anthropic API key, or Codex in API-key mode. Subscription plans that don’t bill per token aren’t what Leash measures.

What happens when my trial or subscription ends?

The 3-day trial needs a card and converts to your chosen plan unless you cancel. If access ends, cloud features become read-only. Your local caps and kill switch keep working.

Your next runaway run stops at your limit.

Start a 3-day trial on Starter or Pro. Card required, cancel any time before it ends. Not ready? Try the demo, no signup.

Protect your next session$ leash run --cap 5 -- claude

Hard caps and a kill switch from day one
Live spend for every session
Starter $11/mo · Pro $22/mo

Start your 3-day trial Try the live demo

Put your agents
on a leash.

Leash sits between your agents and the model provider. It checks every call against your budget, catches runaway loops and gives you a real off switch, before the bill arrives.

Where agent budgets quietly disappear.

How Leash stops a runaway agent.

Your agent makes a call

Claude Code, Codex or your own SDK app sends its request through Leash instead of straight to the provider.

Leash reserves the cost first

Before anything is forwarded, a conservative estimate is reserved against your session, daily and repo budgets.

It spots the loop

Your configured loop and spend-rate rules stop repeated calls and runaway bursts.

You decide what’s next

Get an alert in Slack, Discord or email. Resume, add budget, or keep it stopped.

Put your agents on a leash.

Where agent budgets quietly disappear.

Local keys stay on your machine. You keep control.

Keys stay on your machine

Metadata, not prompts

Limits that work offline

How Leash stops a runaway agent.

Your agent makes a call

Leash reserves the cost first

It spots the loop

You decide what’s next

Three minutes from bug to bill. Leash is in the middle.

Session startswith a $5 cap

Agent hits afailing test

Same fix,eighth attempt

Leash pausesthe session

You resumeor kill it

Four steps. About five minutes.

Install the CLI

Sign in

Set your caps

Run your agent

Your agents, with the controls you need.

Let your agent handle the setup.

See it from your side.

Is Leash right for you?

A great fit if you

Less of a fit if you

A small price for a hard stop.

Starter

Pro

Enterprise

The questions we hear most.

Your next runaway run stops at your limit.

Put your agents on a leash.

Where agent budgets quietly disappear.

Local keys stay on your machine. You keep control.

Keys stay on your machine

Metadata, not prompts

Limits that work offline

How Leash stops a runaway agent.

Your agent makes a call

Leash reserves the cost first

It spots the loop

You decide what’s next

Three minutes from bug to bill. Leash is in the middle.

Session startswith a $5 cap

Agent hits afailing test

Same fix,eighth attempt

Leash pausesthe session

You resumeor kill it

Four steps. About five minutes.

Install the CLI

Sign in

Set your caps

Run your agent

Your agents, with the controls you need.

Let your agent handle the setup.

See it from your side.

Is Leash right for you?

A great fit if you

Less of a fit if you

A small price for a hard stop.

Starter

Pro

Enterprise

The questions we hear most.

Your next runaway run stops at your limit.

Put your agents
on a leash.

Session starts
with a $5 cap

Agent hits a
failing test

Same fix,
eighth attempt

Leash pauses
the session

You resume
or kill it

Put your agents
on a leash.

Session starts
with a $5 cap

Agent hits a
failing test

Same fix,
eighth attempt

Leash pauses
the session

You resume
or kill it