Ratchet

An effect gate for AI agents. Ask before you act; Ratchet answers durably, so the same real-world side effect is attempted at most once, stays inside a declared budget, and leaves an auditable record.

Agents retry. LLM control flow is non-deterministic, network calls fail ambiguously, and processes die mid-action. The result is duplicate emails, double charges, and repeated writes — and nothing in the stack knows which. Vendor idempotency keys help for the few vendors that offer them, and never across separate agent processes or model providers.

Ratchet does not execute your actions. It holds a durable decision record in front of them.

POST /v1/effects/begin  →  decision: execute | duplicate | in_flight
                                    | blocked | approval_required | denied

Only execute authorises the caller to act.

The part that matters

If your process dies between "go" and "done", most systems quietly let the next caller retry. Ratchet won't. The lease expires and the effect becomes indeterminate — a known unknown, surfaced instead of buried. What happens next is the policy you declared for that effect type:

Exactly-once delivery is not achievable in a distributed system and this project does not claim it. What Ratchet guarantees is at-most-once initiation, a recorded outcome that later callers replay, and an explicit state for the case nobody else admits exists.

Related MCP server: OntoRamp Decision Intelligence

Quick start

Requirements: Node 20.11+, Docker (for local Postgres).

npm install
cp .env.example .env          # defaults work for local development
npm run dev:db                # Postgres on :5433 via Docker
npm run migrate
npm run dev                   # control plane on :8787
npm run dev:worker            # lease reaper + webhook delivery (separate terminal)

Then open http://localhost:8787, or drive it from the shell:

bash examples/curl/walkthrough.sh

npm run seed populates a workspace with realistic state — a completed effect, a duplicate, an indeterminate one, and one awaiting approval — so the console has something to show.

Or with Docker Compose

AUTH_SECRET=$(openssl rand -base64 32) docker compose up --build

The core loop

# 1. Ask, before you act.
curl -X POST http://localhost:8787/v1/effects/begin \
  -H "Authorization: Bearer $RATCHET_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "effect_type": "email.send",
    "idempotency_key": "welcome:user_123",
    "payload": { "to": "sam@example.com" },
    "estimated_cost_micros": 800
  }'
# → { "decision": "execute", "effect_id": "eff_...", "lease_token": "lt_..." }

# 2. Do the real thing, yourself. Ratchet never touches it.

# 3. Say what happened.
curl -X POST http://localhost:8787/v1/effects/eff_.../report \
  -H "Authorization: Bearer $RATCHET_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{ "lease_token": "lt_...", "outcome": "succeeded",
        "result": { "message_id": "msg_9f2" } }'

# Any later caller with the same key now gets:
# → { "decision": "duplicate", "result": { "message_id": "msg_9f2" } }

The one rule: report failed only when you know the action did not reach the outside world. If you are unsure — a timeout, a dropped connection — report nothing. The lease lapses and Ratchet records an honest indeterminate. A false failed is worse than silence, because it licenses a duplicate.

Idempotency keys

Derive the key from the work, deterministically.

A key that changes on every attempt makes every retry look like new work.

When Ratchet is unreachable

Ratchet sits in your critical path, so decide this before integrating: on an outage your agent either acts without the gate (fail-open) or refuses to act (fail-closed). Use fail-closed for anything you would have to apologise for; fail-open where the vendor deduplicates anyway.

Full contract, client patterns, and the honest availability posture: docs/FAILURE_MODES.md.

Architecture

                    ┌──────────────────────────────┐
  agents ──────────▶│  control plane (stateless)   │
  REST + MCP        │  Fastify · /v1 · /mcp · web  │
                    └──────────────┬───────────────┘
                                   │
                    ┌──────────────▼───────────────┐
                    │  Postgres                    │
                    │  effects · policies · ledger │
                    │  spend windows · audit       │
                    └──────────────▲───────────────┘
                                   │
                    ┌──────────────┴───────────────┐
  webhooks ◀────────│  worker (long-running)       │
                    │  lease reaper · delivery · GC│
                    └──────────────────────────────┘

The control plane is stateless and scales horizontally — it may run on serverless infrastructure. The worker may not. It expires leases on a timer whether or not a request is in flight; a serverless function cannot do that. Run it as a long-running container. Multiple replicas are safe (every claim uses FOR UPDATE SKIP LOCKED).

At-most-once is enforced by a database unique constraint on (workspace_id, effect_type, idempotency_key) — not by application logic.

Full detail: docs/handoff/ARCHITECTURE.md.

Deploying

The control plane is stateless and can scale freely. The worker cannot — it expires leases on a timer whether or not a request arrives, so it must be a long-running process. That single constraint rules out purely serverless hosts (Vercel, Netlify functions) despite their being easier, and is why fly.toml runs both process groups from one image.

brew install flyctl && fly auth login   # once, needs your browser
npm run deploy:fly

The script is idempotent: it creates the app, provisions managed Postgres, generates AUTH_SECRET once (never rotating it, since that would invalidate every API key), deploys both processes, and verifies readiness. It refuses to proceed unless preflight passes:

npm run deploy:preflight

Preflight runs the full suite and production build, then checks that AUTH_SECRET is strong and not the dev default, PUBLIC_URL is set (otherwise the manifest would advertise localhost), RATE_LIMIT_OVERRIDE is unset, private-network webhooks are off, CORS carries no wildcard, and — if Stripe is selected — that both the key and the webhook secret are present. It prints no secret values.

Any container platform works; only fly.toml is Fly-specific. Set DATABASE_URL, AUTH_SECRET, PUBLIC_URL, NODE_ENV=production, then run node dist/api/server.js (scale freely) and node dist/worker/main.js (at least one, always on).

To rehearse the exact production containers locally:

AUTH_SECRET=$(openssl rand -base64 32) docker compose up --build

Commands

Environment

Every variable, with defaults and safety notes, is in .env.example. The two that are required:

In NODE_ENV=production the process refuses to start if AUTH_SECRET is the development default or shorter than 32 characters, if CORS_ORIGINS contains *, or if WEBHOOK_ALLOW_PRIVATE_NETWORK is on.

For agents

Seven MCP tools: ratchet_begin_effect, ratchet_report_effect, ratchet_check_effect, ratchet_resolve_effect, ratchet_list_effects, ratchet_get_policy, ratchet_usage.

Ready-to-use configs and code: examples/ — Python, TypeScript, curl, Claude Desktop, Cursor, and generic MCP over HTTP.

Security posture

• Only a SHA-256 fingerprint of your payload is stored — never the payload itself.

• API keys are stored as HMAC-SHA256 peppered with a server secret; a database leak alone yields no usable key. Comparison is constant-time and runs even for unknown prefixes.

• Every query is workspace-scoped; a cross-tenant lookup returns 404, never a hint.

• SSRF defence in two layers: static URL validation, then DNS re-resolution with the socket pinned to the checked address on every delivery attempt. Redirects are never followed.

• Agent-supplied text is data, never instructions. Decisions come from stored policy and database state — nothing in a payload can widen a scope, raise a budget, or change a policy.

• Unknown request fields are rejected, not silently dropped.

No third-party audit has been performed. There is no SOC 2 report and no penetration test. Details and threat model: docs/handoff/SECURITY_REVIEW.md.

Pricing

One meter: a gated effect — the first begin for a given (effect_type, idempotency_key). Duplicate suppression, in-flight checks, retries, reports, reads, policy changes, and webhooks are all free. You are never charged for the retry behaviour the product exists to absorb.

Free stops at its allowance unless you load prepaid credit; there is no automatic overage and no invoice. Past roughly 19,300 effects a month, Pro is cheaper than paying credit on Free, so the upgrade is arithmetic rather than a wall.

Two plans, not three: three tiers assert knowledge of three customer segments, and there is no usage history yet to support one. See docs/handoff/PRICING_AND_DISTRIBUTION_REVIEW.md for the reasoning, including why the previous ladder priced a 20x usage range at one number.

Overage draws from prepaid credit, so a runaway agent stops at your balance rather than generating an invoice. Cost model and assumptions: docs/handoff/PRICING_AND_UNIT_ECONOMICS.md.

Payments

Stripe is fully wired: startCheckout creates real Checkout Sessions, and the signed checkout.session.completed webhook credits the ledger. Card details are entered on Stripe's own page and never reach Ratchet, and credit is applied only on the signed webhook — never on the browser returning to a success URL.

Both STRIPE_SECRET_KEY and STRIPE_WEBHOOK_SECRET are required before checkout opens. A key alone selects Stripe but keeps checkout closed: taking a payment that cannot be confirmed would leave a customer charged and uncredited. Run npm run stripe:check to see exactly what is configured (it prints no secret values).

For local development, get a webhook secret without a public URL:

npm run stripe:listen     # prints whsec_… ; put it in .env and restart

Most of Stripe's onboarding checklist does not apply. Ratchet builds line items inline and never reads your Stripe catalog, so there is no product to create; it does not use Stripe Invoicing; and it creates Checkout Sessions through the API, not the no-code builder. What does matter before live payments: verify your account (Stripe's KYC), create a webhook endpoint pointing at your deployed /v1/billing/webhook/stripe and use that endpoint's signing secret in production, and decide whether you need tax collection (STRIPE_AUTOMATIC_TAX, off by default).

With no Stripe credentials at all, the built-in test adapter runs instead: no card is charged and no external request is made. Verified so far in Stripe test mode only — see docs/handoff/KNOWN_LIMITATIONS.md before going live, particularly regarding refunds.

Documentation

License

Apache-2.0.