StitchAPI Docs

Table of Contents

• What is a stitch?

• Motivation

• Why StitchAPI

• What StitchAPI is not

• Features

• Install

• Quick start

• Composition: seam, extends, .with()

• Validation & leveled drift

• Resilience: retry, throttle, timeout

• Caching

• Auth as a boundary

• Surfaces: any request style

• Four front doors

• Agent-native

• Errors & pitfalls

• Zero-infra observability

• Packages

• Documentation

• Contributing

• License

Related MCP server: React Learning MCP Server

What is a stitch?

A stitch is StitchAPI's core primitive: it takes a single endpoint and hands back a callable. You declare the contract once — input, output, auth, resilience — and call it like a local function. The same definition your code calls, the CLI runs and an AI agent invokes — and every caller gets a capability, not a credential. A service with more than one endpoint is a seam — a group of stitches that share a base URL, auth, and one runtime (throttle budget, store, trace sink).

import { stitch } from 'stitchapi';
import { z } from 'zod';

const getUser = stitch({
    path: 'https://demo.stitchapi.dev/users/{id}',
    output: z.object({ id: z.number(), name: z.string() }),
    unwrap: 'data',
});

const user = await getUser({ params: { id: 1 } }); // typed · validated

Keep the fetch or axios you already have — it's the adapter underneath. A stitch sits above the transport and turns an endpoint into a function rather than replacing the call.

Motivation

In almost every project there's a src/api/ folder of thin functions that fire an HTTP request and unwrap the response. Everything that actually makes an integration reliable — auth lifecycle, retries, rate limits, timeouts, response validation, drift detection, observability — gets re-implemented at every call site, and each wrapper rots independently. fetch hands back opaque bytes, and raw bytes aren't what application code (or an AI agent) needs; both want structured, validated, observable results.

A stitch folds all of that back into the call:

Why StitchAPI

There are plenty of ways to get a typed API client — spec-based generators, hand-authored contract clients, workflow platforms, or a folder of hand-rolled fetch wrappers. StitchAPI sits in a spot none of them cover: it turns one endpoint at a time into a resilient, validated, observable function — no spec, no codegen, no config files, no server; only explicit composition.

• Atomic, not spec-first. Spec-based generators (openapi-generator, Orval, Kubb, …) need a complete OpenAPI document first — and most real-world APIs (internal, undocumented, the long tail) never get one. A stitch needs a URL and one example response.

• A runtime, not a code generator. No generated SDK to commit, diff, and regenerate. The declaration is the client, validated on every live call — so a silently renamed field is a loud, leveled drift signal, not an undefined three layers downstream.

• Resilience is declared, not hand-rolled. Retries, throttling, timeouts, circuit breaking, pagination — configuration on the stitch, uniform across every integration.

• Auth is a boundary. A stitch owns its credential and lifecycle; callers get a capability, not the credential. That matters double when the caller is an AI agent.

• Agents are first-class callers. One definition is a function, a CLI command, an HTTP endpoint, and an MCP tool — returning structured, schema-validated, traceable results instead of opaque bytes.

• A library, not a platform. Zero runtime dependencies, embeds in your project, nothing to operate.

For the full competitive landscape and positioning, see the Overview.

What StitchAPI is not

Knowing what a tool refuses to be is how you trust what it is:

• Not an HTTP client or fetch replacement — fetch/axios are the substrate underneath; a stitch sits above the transport.

• Not a code generator — no SDK to commit, diff, and regenerate; the declaration is the runtime, validated live.

• Not spec-first — no OpenAPI document required; a URL and one example response is enough.

• Not a server to deploy — a zero-dependency library you import, for the APIs you don't control.

• Not a workflow engine or iPaaS — no orchestration, queues, or visual builder; composition is plain TypeScript.

• No config files or hidden inheritance — nothing ambient a stitch silently reads.

No server, no codegen, no config files, no implicit inheritance — only explicit composition.

Features

• One primitive, scoped to a surface — stitch(url | config) returns a typed callable; a service with more than one endpoint is a seam that shares base, auth, throttle budget, store, and trace sink across its members.

• Event-stream core — every call yields a typed stream (start → progress → drift → result → done); await is sugar that returns the final validated value.

• Bring-your-own validation — Zod or any Standard Schema validator (Valibot, ArkType, …); types are inferred from the schemas.

• Leveled drift detection — live responses validated against the declared schema (the contract); a required field missing/incompatible throws, while soft drift (a coercion, an undeclared or defaulted field) surfaces as a non-fatal warn / info / verbose finding instead of a silent undefined.

• Declared resilience — retry with backoff and Retry-After, proactive throttle, layered timeouts, a circuit breaker, and idempotency keys.

• Read-through caching — opt-in response cache + in-process coalescing, keyed by a derived, principal-scoped key, loaded lazily from stitchapi/cache.

• Auth as a boundary — bearer, apiKey, basic, cookieSession (auto-login/re-login), oauth2; secrets resolve at call time and never reach the caller.

• Any request style — http by default; graphql, sse, stream, download, llm, shell, and postmessage are peer surfaces behind subpath imports.

• Pluggable state store — throttle counters and sessions behind a 3-method store; swap in Redis/Postgres to go distributed.

• Zero-infra observability — tracing is off by default; opt in per stitch or via STITCH_TRACE_* env vars. No collector, no dashboard.

• Four front doors, one definition — in-process function, CLI (stitch run), HTTP (stitch serve), and MCP (stitch mcp).

• Zero runtime dependencies — "dependencies": {}, built on global fetch, tree-shakeable; ~25 kB min+gzip for the whole entry, ~20 kB for a typical import { stitch }.

Install

npm install stitchapi@rc   # or: pnpm add stitchapi@rc · yarn add stitchapi@rc

Validation is bring-your-own — pass a Zod schema or any Standard Schema validator; none is bundled. The examples below use Zod for familiarity, and hit the live demo API at demo.stitchapi.dev.

Quick start

The smallest stitch is a URL — declare once, call many times:

import { stitch } from 'stitchapi';

const getUsers = stitch('https://demo.stitchapi.dev/users');

const users = await getUsers(); // GET, parsed JSON

Path params use RFC 6570 URI templates; params, query, headers, and body all travel in one input object:

const getUser = stitch('https://demo.stitchapi.dev/users/{id}');

await getUser({ params: { id: 1 }, query: { expand: 'roles' } });
// → GET https://demo.stitchapi.dev/users/1?expand=roles

Reach for the full set of knobs only when you need them — they default off:

const getUser = stitch({
    path: 'https://demo.stitchapi.dev/users/{id}',
    output: User, // a validator of your choice
    unwrap: 'data',
    retry: 3, // ≡ { attempts: 3 }
    timeout: '5s', // ≡ { total: '5s' }
    cache: '1m', // ≡ { ttl: '1m' }
});

const user = await getUser({ params: { id: 42 } });
// → typed · validated · retried · cached

A stitch yields a typed event stream; await consumes it and returns the final value, while .safe() resolves to { ok, data, error } and .stream() yields progress, throttle waits, retries, and drift as they happen.

Composition: seam, extends, .with()

Everything reusable is a named value, and a stitch composes values — no global config is ever required. A service with more than one endpoint is a seam: declare the shared base, auth, and budget once; each endpoint is a member that inherits the config and shares one runtime (throttle bucket, store, trace sink):

import { seam } from 'stitchapi';

const api = seam({
    baseUrl: 'https://demo.stitchapi.dev',
    retry: { attempts: 3, on: [429, 503] },
    timeout: { total: '30s' },
});

const listUsers = api.stitch({
    path: '/users',
    output: User.array(),
    unwrap: 'data',
});
const getUser = api.stitch({
    path: '/users/{id}',
    output: User,
    unwrap: 'data',
});

For lighter reuse, extends: [fragment | stitch] merges config left→right (own fields win), and .with() pre-binds part of the input while reusing the same runtime. Full guide: Authoring.

Validation & leveled drift

The declared output schema is the contract. Wrap it in drift(): a response is validated (the call returns the validated value — coerced, defaulted, unknown keys stripped), and the difference between the raw body and that validated value is reported as a leveled, non-fatal signal:

import { drift, stitch } from 'stitchapi';
import { z } from 'zod';

const listOrders = stitch({
    path: 'https://demo.stitchapi.dev/users/{id}/orders',
    unwrap: 'data',
    output: drift(
        z.array(z.object({ id: z.number(), total: z.number().optional() })),
        {
            ignore: ['[].meta'], // acknowledged, unconsumed — don't report it
            severity: { coerced: 'info' }, // re-level a kind, or pass a level/list to filter
        },
    ),
});

Drift is schema-anchored — no snapshot to manage. Severity lives in the schema: a required field missing/incompatible is a hard invalid that throws; everything else is non-fatal drift on the event stream. Declared variance (an optional field, a nullable, an empty array) validates clean, so it's never a false alarm; ignore silences known-but-unconsumed fields and severity filters or re-levels the soft signals. The request side validates too: input takes a schema per part and fails fast before any request is sent. Full guide: Validation & drift.

Resilience: retry, throttle, timeout

The things every src/api/ folder reinvents are configuration here — uniform across every stitch:

const listUsers = stitch({
    baseUrl: 'https://demo.stitchapi.dev',
    path: '/users',
    retry: { attempts: 4, on: [429, 502, 503], respectRetryAfter: true },
    throttle: { rate: '1/s', concurrency: 2, scope: 'host' },
    timeout: { total: '30s', perAttempt: '10s' },
});

throttle is proactive (keeps you under a limit before it bites; scope: 'host' shares a limiter across stitches), retry is reactive (backoff + Retry-After), and timeout aborts with a real AbortSignal. Three more knobs round it out: circuit fast-fails a dependency that's already down, idempotency injects a stable Idempotency-Key on writes, and acceptStatus treats a non-2xx (e.g. 404) as a normal result instead of a throw. Full guide: Resilience.

Caching

A read-through response cache with in-process request coalescing — off by default, loaded lazily from stitchapi/cache only when a stitch sets cache. The key is derived from the resolved request (no caller-authored keys to drift) and principal-scoped by default, so user A is never served user B's cached response:

const getUser = stitch({
    path: 'https://demo.stitchapi.dev/users/{id}',
    output: User,
    unwrap: 'data',
    cache: '5m',
});

const listAnnouncements = stitch({
    path: 'https://demo.stitchapi.dev/announcements',
    output: z.array(z.object({ id: z.number(), title: z.string() })),
    unwrap: 'data',
    cache: {
        ttl: '1h',
        scope: 'app',
        vary: ['accept-language'],
        maxEntries: 500,
        version: 1, // pins the shape — cacheable without a fingerprinter
    },
});

Caching is sound by construction: a stitch with an output schema caches only when that schema can be fingerprinted or you pin a version — otherwise it refuses to cache rather than serve a stale shape. Mark sensitive data sensitive: true to opt out entirely.

Auth as a boundary

Auth is a field on the stitch — never global. Secrets resolve at call time (env(), secretsFile()), the declaration is committable, and the caller gets data without ever seeing the credential:

import { bearer, env, stitch } from 'stitchapi';

const getUser = stitch({
    path: 'https://demo.stitchapi.dev/users/{id}',
    auth: bearer(env('API_TOKEN')), // resolved per call; the caller passes no secret
});

bearer, apiKey, and basic are header strategies; oauth2() runs the client-credentials grant and caches/refreshes the token; and cookieSession runs a login, captures the cookie, replays it, and re-logs-in when the wall returns — the caller never sees the password or writes the cookie dance. Full guide: Auth.

Surfaces: any request style

A surface is the request style a stitch speaks. http is the default; the rest are peer surfaces on the same engine — auth, retry, throttle, timeout, validation, and the event stream compose with every one. Each ships as its own subpath import, so import { stitch } pulls in http alone.

Full guide: Surfaces.

Four front doors

The same typed unit is reachable four ways, so humans and agents call exactly the same validated, observable thing:

stitch run streams every event as one JSON line on stdout (ready for jq); stitch trace summarizes the run log — runs, failures, retries, drift, and latency percentiles. Full guide: Surfaces → CLI.

Agent-native

A stitch is built to be invoked by an AI agent. Auth lives on the stitch, so an agent gets a capability, not a credential. And rather than one MCP tool per endpoint (which floods the model's context), an agent drives a single code-mode tool, run_stitch: it writes a small TypeScript snippet that imports stitchapi and calls the stitch, with list_stitches and describe_stitch for discovery — so the context budget stays flat as the catalog grows.

Authoring is agent-friendly too — hand an agent one curl/HAR/doc example and it emits a stitch declaration, with a deterministic shortcut for the common case:

$ stitch from-curl 'curl https://demo.stitchapi.dev/users/7 -H "authorization: Bearer …"'
# prints a ready-to-commit stitch: id-like segments lifted to {params}, secrets → env()

stitch init writes the “declare a stitch, don't hand-roll fetch” rule into the files coding agents read (AGENTS.md, Cursor/Windsurf/Cline rules, a CLAUDE.md section), and the docs build emits an auto-generated llms.txt. Full guide: Use from an agent.

Errors & pitfalls

A failed stitch throws a StitchError — an Error subclass carrying .status, .attempts, and (for response failures) .body (the parsed error payload, on a non-enumerable channel that never reaches a trace sink) and .url. Prefer branching? .safe() resolves to { ok, data, error }. Each failure mode has a stable code and a docs page:

Full catalog: Errors & pitfalls.

Zero-infra observability

Observability is a consumer of the event stream, not a separate system — and it's off by default. Opt in per stitch (trace: 'console' / fileSink(path) / a TraceSink) or globally with env vars — no collector, no dashboard, nothing to deploy:

STITCH_TRACE_CONSOLE=1 node app.js          # live, colored, one line per event on stderr
STITCH_TRACE_FILE=./run.jsonl node app.js   # append JSONL to a path
STITCH_EXPORT=otlp node app.js              # also fan events to an OTLP collector

Built-in sinks scrub secrets at the sink boundary (the live request is never touched). stitch trace summarizes the JSONL; @stitchapi/pino ships the same events as structured logs.

Packages

This repository is a pnpm workspace. The published library is stitchapi; the rest are thin, peer-dependency integrations that add no capability of their own. The table below is generated by pnpm gen:readme (and verified in CI by pnpm check:readme): it lists every publishable workspace package, so it never drifts as packages are added. Each package's own README links to it on npm.

Documentation

The full documentation site lives at stitchapi.dev — Quickstart, per-feature Guides, Concepts, Surfaces, For agents, and the generated Reference.

The published library's own README (what npm renders) is in packages/core/README.md. Design notes live in docs/: Feature Lenses, Overview, Design.

Contributing

Local setup, the dev loop, the bundle-size budget, and the worktree + PR workflow are documented in CONTRIBUTING.md.

License

Apache-2.0