CommitLayer documentation
Everything needed to record what an agent does, supply the facts from your systems of record, and read the verdict. The quickstart takes about ten minutes against the free Developer plan; the reference pages are exported from the API repository and carry the date they last changed.
Guides
Quickstart
From an API key to a verified refund in five steps: two keys, the first-run contract scop_refund_order, record the action, push the facts, read the verdict.
After the first verdict
Dry runs, webhooks, the completion gate, settlement, the SDK decorator, drafting a contract from what was recorded, MCP, OpenTelemetry and orphan effects.
Beyond one action
Three checks that look across actions or across time: a total per agent over a window, a value compared with the request, and a later record that undoes the action.
Reference
Every endpoint, the contract language, the MCP server, the OpenTelemetry convention and the sandbox.
- API reference
Every endpoint of the HTTP API: request and response shapes, error codes, limits, idempotency, dry runs, and what each plan allows.
- Action contracts
The YAML that states what must hold before an agent acts, the limits on the action, and the effects that must and must not appear in the systems of record afterwards.
- MCP server and connectors directory
The MCP server assistants read verdicts and evidence from, the OAuth flow, the tool list with its annotations, and the Claude connectors directory listing.
- OpenTelemetry convention
The span attributes that turn an agent runtime's OpenTelemetry traces into outcome claims and runtime-context facts, with the OTLP endpoint to send them to.
- The sandbox and the failure matrix
How CommitLayer is tested without customers: a small enterprise of fake systems, a scripted or model-driven agent, injected failures, and an oracle that scores every verdict.
Workflow packs
Contracts, evidence rules, recorder mappings and test scenarios for Stripe customer operations, x402 service delivery and Coinbase agent trades, and how evidence provenance is enforced.
- Stripe Customer Operations pack
Five workflows for customer-operations agents (refund an order, cancel a subscription, issue a credit note, adjust an invoice, resolve a case) verified against Stripe, Shopify and the help desk; installs as drafts, a person activates.
- Stripe pack: refund an order
The refund_order workflow: one Stripe refund for the approved amount, the order cancelled, the case resolved, the customer told; every check, its parameters and what each failure looks like.
- Stripe pack: cancel a subscription
The cancel_subscription workflow: the subscription cancelled in Stripe with the approved timing, the case resolved, the customer told; every check, its parameters and what each failure looks like.
- Stripe pack: issue a credit note
The issue_credit_note workflow: one credit note on a paid invoice for the approved amount, no cash refund alongside; every check, its parameters and what each failure looks like.
- Stripe pack: adjust an invoice
The adjust_invoice workflow: an open invoice reduced once by the approved amount and the customer told; every check, its parameters and what each failure looks like.
- Stripe pack: resolve a customer case
The resolve_customer_case workflow: the case solved in the help desk and the customer told within a day; every check, its parameters and what each failure looks like.
- Stripe pack: connector requirements
Which connectors the Stripe pack needs, the read permissions to grant on each, what is read through it, and what to push instead when a system has no connector.
- Stripe pack: recorder and OpenTelemetry setup
Running the MCP recorder with the Stripe pack's mapping, changing the mapping, the warning for tools without a rule, and what OpenTelemetry ingestion adds.
- Stripe pack: REST, Python, TypeScript and MCP examples
The pack's REST, Python and TypeScript examples: record an action, push the facts only your systems hold, read the verdict.
- Stripe pack: from a test-mode key to a first verdict
From a Stripe sandbox's restricted key to a first verdict, command by command: connect, install, activate, act, push one fact, read PASS, then break it once.
- Stripe pack: reading a verdict and its reason codes
Every verdict and reason code the Stripe pack can return, which workflow reports it, what it means and what to do about it.
- x402 Service Delivery pack
Nine workflows for x402: six purchases, the provider's side of a paid request, and prepaid balances (a top-up and the calls that draw on it); the payment settled once for the quote and the service was delivered, two evidence paths, a pending settlement UNKNOWN.
- Coinbase Agent Trade pack
Trades, closes, rebalances and payments an agent makes on a Coinbase account, verified against Coinbase's orders, fills, balances and sends, and onchain swaps from a CDP Server Wallet verified against the chain; version 2 of coinbase_agent_trade as an upgrade beside version 1.
- Evidence provenance
The trust class every fact carries, which key class may push what, and why an agent's own report never satisfies a pack contract.
Connectors
What each read-only connector reads, with which permissions, what it never touches, and the facts it produces.
- Stripe
What the Stripe connector reads, with which permissions, what it never touches, how it syncs, and the facts it produces.
- Shopify
What the Shopify connector reads, with which permissions, what it never touches, how it syncs, and the facts it produces.
- Salesforce
What the Salesforce connector reads, with which permissions, what it never touches, how it syncs, and the facts it produces.
- HubSpot
What the HubSpot connector reads, with which permissions, what it never touches, how it syncs, and the facts it produces.
- Zendesk
What the Zendesk connector reads, with which permissions, what it never touches, how it syncs, and the facts it produces.
- Intercom
What the Intercom connector reads, with which permissions, what it never touches, how it syncs, and the facts it produces.
- Coinbase
What the Coinbase connector reads, with which permissions, what it never touches, how it syncs, and the facts it produces.
- EVM chain (JSON-RPC)
What the EVM chain (JSON-RPC) connector reads, with which permissions, what it never touches, how it syncs, and the facts it produces.
- CSV, XLSX and Parquet upload
What the CSV, XLSX and Parquet upload connector reads, with which permissions, what it never touches, how it syncs, and the facts it produces.
- OpenTelemetry ingestion
What the OpenTelemetry ingestion connector reads, with which permissions, what it never touches, how it syncs, and the facts it produces.
SDKs
Record actions, push facts and read verdicts from Python, TypeScript, Go and Java.
- Python SDK
Record agent actions, push facts from systems of record, read verdicts and draft contracts. Standard library only; Python 3.10+.
- TypeScript SDK
Record agent actions, push facts from systems of record, read verdicts and draft contracts. Zero runtime dependencies; Node 20+ (uses the global fetch and crypto). Ships with TypeScript types.
- Go SDK
Record agent actions, push facts from systems of record, read verdicts and draft contracts. Standard library only; Go 1.22+.
- Java SDK
Record agent actions, push facts from systems of record, read verdicts and draft contracts. The runtime has zero dependencies (JDK HTTP client, its own JSON codec); Java 17+.
MCP recorder
A stdio proxy that records a mapped MCP tool call as an action before forwarding it.
- MCP recorder (Python)
A small stdio proxy that sits between an MCP client (Claude Desktop, Claude Code, Cursor, an agent framework) and any downstream MCP server (a Stripe or Zendesk MCP server, your own). It forwards every JSON-RPC…
- MCP recorder (TypeScript)
A small stdio proxy that sits between an MCP client (Claude Desktop, Claude Code, Cursor, an agent framework) and any downstream MCP server (a Stripe server, a Zendesk server, your own). It forwards every JSON-RPC…
Agent frameworks
Record an agent's tool calls before they run, in LangGraph, CrewAI, the OpenAI Agents SDK and Pydantic AI.
- LangGraph and LangChain agents
Record a tool call as a CommitLayer action before the tool runs, then gate the graph on the verdict.
- CrewAI
Record a crew's tool calls as CommitLayer actions before the tools run.
- OpenAI Agents SDK
Record function-tool calls as CommitLayer actions before they run.
- Pydantic AI
Record tool calls as CommitLayer actions before the tools run.
Workflow runtimes
Record before the activity or step, wait on the verdict after, in Temporal and Inngest.
- Temporal and Inngest (Python)
CommitLayer for agent frameworks and durable-workflow runtimes. One distribution, six optional packages, built on the commitlayer SDK:
- Temporal and Inngest (TypeScript)
CommitLayer for workflow engines. Two subpaths, one package, zero runtime dependencies beyond the engine you already run (@temporalio/ and inngest are optional peer dependencies; the matching SDK is loaded lazily, so…
- Temporal (Python)
CommitLayer for the Temporal Python SDK. Temporal makes the workflow durable. CommitLayer independently verifies the result.
- Inngest (Python)
CommitLayer for the Inngest Python SDK. The same positioning as the Temporal package, in Inngest's terms: Temporal makes the workflow durable. CommitLayer independently verifies the result. Here Inngest makes the…
Releases
Every page here has a markdown version, listed in /llms.txt for the tools that read documentation as text. Questions: ask us.