Introduction

Confirm is the approval layer for AI agents. Your code calls one API before an agent does anything sensitive. The agent pauses, a human decides, and a signed webhook resumes the run. Every decision lands in an immutable audit log.

Why this exists

Agents are good enough to act and not good enough to act unsupervised. The failure modes are concrete: a support agent refunds $10,000 instead of $10, a sales agent invents a fact in a cold email, an ops agent drops a table that billing still reads. Teams either ship agents with no brakes, or bolt approvals onto Slack threads that satisfy no auditor.

Confirm makes the brakes infrastructure: one API to pause, one review page for the human, one webhook to resume, one log for compliance.

The request lifecycle

Create. Your code calls POST /v1/requests with a summary, the exact action payload, and who decides. That can be a single email or an approver group like "group:finance" for escalation.
Notify. Every approver gets their own single-purpose magic link by email. No account, no app. Optionally, a Slack channel notification announces the request to the team.
Decide. The review page shows the payload plus the context you passed: reasoning (why the agent wants this) and recentActions (what it did first), so the decision takes seconds. The human approves, rejects, or edits the payload before approving. Identity, timestamp, IP, and note are recorded.
Resume. We fire an HMAC-signed webhook carrying the verdict, the effectivePayload (human edits win), and any agentState you parked with the request, so a fresh worker can rehydrate and continue even hours later.
Expire. If nobody answers within the TTL, the request becomes EXPIRED and a webhook tells your system. Silence never becomes consent.

How you integrate

Three ways in, same approval underneath. Pick by stack.

Path	Best for	Looks like
TypeScript SDK `@confirm/sdk`	Node/TS agents. The recommended path.	`guard()` gates your whole tool layer with policies (fail-closed), or `withApproval()` wraps one function.
MCP server `confirm-mcp`	MCP agents: Claude Desktop, Cursor.	Adds a `request_approval` tool with one config block, no code.
Raw HTTP	Any language, any framework.	One `POST` with a bearer key. Resume on the webhook, or poll.

What ships in the box

Capability	What it does
Escalation routing	`notify: "group:finance"` fans out to a named approver group. First decision wins, and the audit names exactly who decided.
Approver context	`reasoning` and `recentActions` render the agent's story above the payload on the review page.
Payload editing	Approvers fix the JSON before approving. Your code executes `effectivePayload`, never the first draft.
State parking	`agentState` stores up to 256KB of serialized context and returns it in the webhook for rehydration.
Fail-safe expiry	Workspace default or per-request TTL. Unanswered requests expire and notify your agent.
Signed webhooks	HMAC-SHA256 over a timestamped payload, retried, with delivery receipts.
Teams	Invite collaborators with roles. Approvers never need accounts.
Audit log	Append-only record of every request, decision, and delivery attempt.

Request statuses

Status	Meaning
PENDING	Waiting on a human. The agent should stay paused.
APPROVED	A human said yes. Execute `effectivePayload`.
REJECTED	A human said no. Do not execute; the rejection note explains why.
EXPIRED	The TTL elapsed with no decision. Treat it as a rejection.

PENDING is the only non-terminal state. Terminal states never change: a request can't be re-approved, un-rejected, or resurrected after expiry.

The quickstart takes you from zero to a resolved approval in about five minutes with the SDK. Building in a framework? Jump to the integration guides.