GovernanceforAIagents.Onelineofcode.

Register every agent, scope what it can touch, watch what it does in real time, and kill it mid-run when it goes sideways.

Works with
ClaudeGPTCursorClaude CodeLangChainMCP

See it in action

Watch Leash govern an agent through three real scenarios — a clean run, an approval gate, and a kill switch.

leash dashboard — refactor-bot
Clean Run/
Approval Flow/
Kill Switch
Agent Activity
Governance
Status
Running
Risk Score8
Scope Checks
agent:refactor-bot • session:a3f8c2
scenario 1/3

Your AI agents have admin access and nobody's watching.

Every AI agent you deploy acts as a super-admin by default. It has the same access as the service account it runs on — and most of those accounts have far more permissions than any agent needs.

80% of organizations have already experienced unintended agent behavior. Agents that read files they shouldn't, call APIs outside their scope, or make decisions no human approved.

Existing identity tools were built for humans who log in once a day. They weren't designed for agents that make 5,000 API calls a minute, spawn sub-agents, and operate autonomously around the clock.

YOURINFRACoding AgentAISupport BotAIData PipelineAIRogue Agent⚠Research BotAIBLOCKED

Everything you need to govern your AI agents

From registration to revocation, Leash gives you complete visibility and control over every agent in your stack.

Agent registry

Every agent gets an ID, a human owner, and declared scopes. One SDK call to register, no IT ticket required.

Scoped permissions

Least-privilege access enforced on every tool call. Agents only touch what you explicitly allow.

Live audit log

Every action, every decision, every risk score — searchable, exportable, and ready for your next compliance review.

Anomaly detection

Rule-based plus LLM-powered risk scoring flags suspicious action sequences before they cause damage.

Human approval flow

Pause high-risk actions automatically. Approvers get pinged on Slack, SMS, or email. One tap to allow or deny.

Kill switch

Revoke an agent mid-execution. JWT invalidated, sessions terminated, pending actions cancelled in under five seconds.

Three lines of code. Complete control.

1

Register your agent

Give it an ID, assign a human owner, and declare exactly what it's allowed to touch.

2

Wrap your runtime

One function call wraps your existing agent. No refactoring, no migration.

3

Watch the dashboard

Every action streams to your audit log in real time. Set alerts, require approvals, or pull the kill switch.

agent.ts
1import { Leash } from "@leash-ai/sdk"
2
3const leash = new Leash({ apiKey: process.env.LEASH_KEY })
4
5// 1. Register the agent
6const agent = await leash.agents.register({
7  name: "refactor-bot",
8  owner: "eng-team@acme.com",
9  scopes: ["repo:read", "repo:write", "ci:trigger"],
10})
11
12// 2. Wrap your runtime
13const session = await leash.sessions.start({
14  agentId: agent.id,
15  ttl: "1h",
16  approvalPolicy: "auto",
17})
18
19// 3. Every action is now logged
20await session.action("file.read", {
21  path: "/src/index.ts",
22})
Live interactive demo

See the dashboard in action.

Run three real scenarios — a clean execution, an approval gate, and a kill switch — in a fully interactive dashboard.

Launch the demo→

Get early access.

Founding customers get six months free and a direct line to the roadmap. Put your agents on a leash before they put you on notice.