@usewrit/agent-sdk

Official TypeScript SDK for the Writ local agent (writ-agentd) — the loopback HTTP API (http://127.0.0.1:8131/v1/…) served by the Writ desktop app and the self-hosted daemon.

• Zero runtime dependencies (Node ≥ 18, global fetch)
• ESM with full type declarations, TypeScript strict mode
• Automatic daemon discovery (~/.writ/runtime.json), typed errors, uniform list pages, typed SSE run events, runAndWait

Install

npm install @usewrit/agent-sdk

Quickstart

import { WritAgent, runRowId } from "@usewrit/agent-sdk";

const client = new WritAgent(); // discovers the running agent + token
const { data: workflows } = await client.workflows.list();
const run = await client.workflows.runAndWait(workflows[0].id, { inputs: { city: "Paris" } });
const { data: rows } = await client.runs.data(runRowId(run));
console.log(run.status, rows);

(Run-feed ids are composite strings like "workflow-3"; the exported runRowId(run) helper extracts the numeric row id that runs.get/data/cancel/events take.)

Auth & discovery

Every request carries Authorization: Bearer <token>. The token families are wlt_ (runtime token, full access), wlk_ (scoped API key minted via client.keys.create), and wlo_ (OAuth access token) — the SDK treats them as opaque strings.

If you pass nothing, the client discovers a live daemon on first use (single-flight, cached for the client's lifetime):

1. Env overrides — WRIT_API_URL (base URL) and WRIT_TOKEN (bearer). Both set → discovery is done. One set → it fills that field, the rest is discovered.
2. runtime.json descriptors, in order, first live one wins:
   1. $WRIT_HOME/runtime.json
   2. ~/.writ/active_profile → ~/.writ/profiles/<id>/runtime.json
   3. ~/.writ/runtime.json
   4. every ~/.writ/profiles/<id>/runtime.json (capped, deduped)
3. Each candidate is liveness-probed (GET /v1/agent, 2 s budget); a stale descriptor from a crashed daemon falls through to the next candidate.

Explicit options always win over discovery:

const client = new WritAgent({
  baseUrl: "http://127.0.0.1:8131", // no trailing /v1
  token: process.env.MY_WRIT_KEY!,  // e.g. a scoped wlk_ key
  timeout: 30_000,                  // per-request, ms (default 30 s)
});

Discovery is filesystem-based and desktop-only. In a browser (or any non-Node runtime) you must pass baseUrl + token explicitly — otherwise construction of the first request throws a WritDiscoveryError telling you so. For the HTTPS twin port (https://127.0.0.1:8132, local CA at ~/.writ/tls/ca.pem), pass a custom fetch wired to your TLS setup via the fetch option.

Prefer failing fast? await WritAgent.discover() (or client.connect()) resolves and probes eagerly instead of on the first call.

Errors

Three kinds, one base class — no automatic retries in v1:

import { WritApiError, WritConnectionError, WritDiscoveryError } from "@usewrit/agent-sdk";

try {
  await client.workflows.get(999999);
} catch (err) {
  if (err instanceof WritApiError) {
    // err.status === 404, err.code === "not_found",
    // err.message === "not found: workflow 999999", err.body = parsed JSON
  } else if (err instanceof WritConnectionError) {
    // daemon went away mid-session / request timed out
  } else if (err instanceof WritDiscoveryError) {
    // no live daemon found — is the Writ agent running? set WRIT_TOKEN, or pass token
  }
}

Domain errors are JSON {"error", "code"}; plain-text rejections (e.g. axum body-deserialize) map to a status-derived code (400→bad_request, 404→not_found, 409→conflict, 422→unprocessable, 5xx→internal, …).

One deliberate non-error: runs.cancel(id) / workflows.cancel(id) return the daemon's 409 {status: "not_running"} as a result, not an exception — an already-finished run is a valid answer to "cancel".

Lists are always a Page

The daemon mixes three list envelopes; every SDK list method normalizes them:

const page = await client.runs.list({ limit: 20, status: "failed" });
page.data;  // T[]
page.count; // items in this page
page.total; // number | null (only the runs feed reports a total)

Run events (SSE)

runs.events(runId) is a typed async iterator over the daemon's text/event-stream (started | step | progress | finished | error); it completes after the terminal event. Keep-alive comments are handled for you.

const { run_id } = await client.workflows.run(3, { inputs: { q: "writ" } });
for await (const ev of client.runs.events(run_id)) {
  if (ev.event === "step") console.log(`step ${ev.index} (${ev.step_type}): ${ev.status}`);
  if (ev.event === "finished") console.log("done:", ev.status);
  if (ev.event === "error") console.error("failed:", ev.message);
}

workflows.runAndWait(id, opts) wraps the whole dance: start the run, follow the SSE stream, fall back to 1 s polling if the stream drops, and return the final run (optionally with includeResults: true). The default deadline is 600 s (waitTimeout, ms); on expiry it throws — the run itself is not cancelled (call runs.cancel(runId) if you want that).

Waiting: three ways

The daemon's run endpoint is async by default and you choose how to wait:

| You want | Call | You get | | --- | --- | --- | | A handle to watch yourself | workflows.run(id) | 202 {run_id} — stream runs.events(run_id) | | Just the answer, one request | workflows.run(id, { wait: true }) | The terminal run document | | Live events + the enriched feed item | workflows.runAndWait(id) | Final RunFeedItem (+ results) |

wait: true is the server-side wait (?wait=true) — the daemon blocks and answers with the run's own result. No SSE, no poll loop:

const done = await client.workflows.run(3, { wait: true, timeout: 60 });
if (done.status === "success") console.log(done.data);
else console.error(done.error);          // a FAILED run resolves — check status

A run that fails is a result, not an exception. Only an expired budget rejects, as a WritRunTimeoutError carrying the still-valid run id — the run keeps going, so collect it rather than retrying (a retry starts a second run):

import { WritRunTimeoutError } from "@usewrit/agent-sdk";

try {
  await client.workflows.run(3, { wait: true, timeout: 30 });
} catch (err) {
  if (err instanceof WritRunTimeoutError) {
    for await (const ev of client.runs.events(err.runId)) { /* … */ }
  }
}

Reach for runAndWait when you want live progress or a deadline longer than the daemon's own ceiling (timeout is clamped server-side to 1–3600 s, default 120).

Surface

| Namespace | Endpoints | | -------------------- | ------------------------------------------------------------------------- | | client.agent | status(), health() | | client.workflows | list, create, get, update, delete, run, cancel, session, clearSession, runAndWait | | client.runs | list, get, results, data, dataCsv, events, cancel | | client.monitors | list, create, get, update, delete, run, changes, capacity, recentChanges | | client.selectors | list, create, get, update, delete, toggle, test, setBaseline, clearBaseline (under a monitor) | | client.extractors | list, create, get, update, delete, toggle, test | | client.automations | list, create, get, update, delete, enable, run | | client.personas | list, create, get, update, delete, runs, validateTotp, test2fa | | client.secrets | list, set, get, delete — metadata only; values never come back | | client.vault | status, lock, unlock | | client.files | list, upload (multipart), fromData, get, delete, content (bytes) | | client.data | query, workflowData, deleteWorkflowData, facets, export, dataRuns | | client.crawl | list, start, get, cancel — the "Dragnet" whole-site crawl (local daemon) | | client.cloud | scrape, map, crawl, crawlStatus, quota — the tiered cloud surface | | client.keys | list, create (plaintext key shown once), get, delete — needs wlt_ | | client.wsTicket | mint a single-use WS ticket (record / ai-preview) |

Cloud tier (client.cloud) — scrape / map / crawl

These three verbs run on Writ Cloud, never on the calling machine, with a Firecrawl-style tier resolved from your credential:

// Metered: pass an API key (or set WRIT_API_KEY) → billed per page, uncapped by your plan.
const client = new WritAgent({ apiKey: "wt_…" });
const page = await client.cloud.scrape("https://example.com");   // { markdown, tier: "metered", … }
const job  = await client.cloud.crawl({ url: "https://example.com" });

// Keyless: no key → free, daily-capped per install (X-Writ-Client-Id) and per IP. scrape + map only.
const anon = new WritAgent();
const p = await anon.cloud.scrape("https://example.com");        // { markdown, tier: "keyless", quota }
await anon.cloud.crawl({ url: "…" });                            // throws WritApiKeyRequiredError

Credential precedence: apiKey → WRIT_API_KEY → keyless. Errors: 429 → WritRateLimitedError (resetAt), 402 api_key_required → WritApiKeyRequiredError, other 402 → WritInsufficientCreditsError.

Secret hygiene mirrors the daemon: workflow credentials, persona passwords and vault secret values are sealed daemon-side and are never present in any response this SDK returns.

The full wire contract lives in ../openapi/writ-agent.yaml and the cross-language rules in ../DESIGN.md.

Roadmap (not in v1)

OAuth AS endpoints, /mcp JSON-RPC, the OpenAI-compat routes, cloud/* surfaces, backup/update/network/tls/settings/diagnostics/notifications, AI surfaces (ai-assist / ai-sessions / ai-concierge / streaming sessions), relay, hooks, and a WebSocket convenience for opening the recorder socket (the SDK mints the ticket; opening the socket is up to you for now).

Development

npm install
npm run build     # tsc → dist/
npm test          # vitest against an in-process mock daemon
WRIT_SMOKE=1 npm test   # additionally smoke-test against a real, running agent

MIT © Writ