Skip to content

api.md

Latest commit

 

History

History
157 lines (125 loc) · 6.62 KB

api.md

File metadata and controls

157 lines (125 loc) · 6.62 KB
Raw
description Use the authenticated JSON API, published OpenAPI document, and typed TypeScript client.

API & client

marimohub exposes a JSON HTTP API under /api/v1/* and ships a typed TypeScript client generated from its OpenAPI document.

Response envelope

JSON responses under /api/v1/* use one envelope:

// success
{ "success": true, "data": { /* … */ } }
// failure
{ "success": false, "error": { "code": "FORBIDDEN", "message": "…" } }

Content routes return raw bytes on success. These include notebook HTML snapshots, workspace files and archives, and job output and logs. Their errors still use the JSON envelope.

Authentication is via the session cookie issued by your auth backend, or a personal access token sent as Authorization: Bearer … (for CI, scripts, and the CLI). Project reads require effective viewer access through ownership, membership, or MARIMOHUB_DEFAULT_ROLE; none hides non-member projects. Writes are role-gated. A project audit log requires project manager. The deployment audit log requires a super admin (see Security → Authorization). Editor ownership, temporary session creation, and takeover are documented in Editor sessions.

Scoped tokens further restrict the actions and projects available to automation. Security labels can deny access even when the project role permits it. Neither mechanism grants extra project permissions.

Endpoints

The docs site publishes the OpenAPI 3.1 document at /openapi.yaml from the same source checkout used to build these docs. Use that URL for code generation and offline tooling.

A running hub also serves GET /api/v1/doc. That endpoint is protected like the rest of /api/v1/*, so send a session cookie or PAT:

export MARIMOHUB_URL=https://hub.example.com
export MARIMOHUB_TOKEN=mhub_pat_…
curl --fail --location \
  --header "Authorization: Bearer ${MARIMOHUB_TOKEN}" \
  "${MARIMOHUB_URL}/api/v1/doc" \
  --output openapi.yaml

The repository source is packages/api/openapi.yaml.

Resource groups:

  • Projects — list/create/update/delete projects; add/update/remove members (/projects/{pid}/members). Project responses carry your_role (the caller's effective role). Managers can read the audit log one UTC day at a time (GET /projects/{pid}/events?date=YYYY-MM-DD, defaults to today) — every project/notebook mutation is recorded as an event.
  • Audit — super admins can read deployment events with GET /events. The endpoint returns newest events first. It supports exact filters for event type, actor ID, and project ID. The default range is the last 30 UTC days. A custom inclusive range cannot contain more than 30 days.
  • Notebooks — create and manage local or Git-synced notebooks, read code, manage versions, and rotate notebook sync tokens.
  • Sessions — list, create, inspect, heartbeat, and stop kernel sessions. The session routes also expose editor ownership and exclusive takeover. Secondary surfaces provide VS Code and OpenCode access within edit sessions.
  • Workspace files — browse, read, upload, copy, move, and delete notebook files under /projects/{pid}/notebooks/{nid}/workspace.
  • Jobs — manage job definitions, trigger or cancel runs, and read run history, HTML output, and logs under /projects/{pid}/notebooks/{nid}/jobs.
  • Integrations — discover integration kinds and manage project or organization integration instances. Each kind reports its available secret sources. Version-history lists use pagination.
  • Users and tokens — resolve or search users, and create, list, or revoke personal access tokens.
  • System — GET /api/v1/version and GET /api/v1/capabilities report deployment information. GET /api/health is the unversioned health probe.

The MCP server uses /mcp with separate OAuth discovery and authorization endpoints. These endpoints follow MCP and OAuth protocols rather than the JSON API envelope.

Pagination

The project, notebook, notebook-version, project-session, integration-instance, integration-version, job, job-run, and deployment-audit list endpoints return this page shape:

{
	"success": true,
	"data": {
		"items": [
			/* … */
		],
		"next_cursor": "MTAw",
	},
}

Pass ?limit= to set the page size. Pass a prior next_cursor as ?cursor= to get the next page. Items are ordered newest-first. A next_cursor value of null marks the final page. The cursor is opaque.

Some small or naturally bounded collections still return arrays. These include project members, API tokens, integration kinds, project daily audit events, and user search results. The OpenAPI response schema is authoritative for each route. GET /api/v1/capabilities reports the default and maximum page sizes and other server limits.

Caching

Most successful reads carry an ETag and Cache-Control: no-cache. Send the ETag back as If-None-Match to revalidate; an unchanged resource answers 304 Not Modified with no body. Browsers do this automatically, which keeps the session-status poll loop cheap.

Content and credential routes can set stricter cache policies, including Cache-Control: private, no-store. Use the response headers for each route.

Integration updates

An integration configuration is one versioned resource. The API does not provide per-secret update routes. Get the integration and keep its ETag before you change its configuration. Send that ETag as If-Match with the write.

If another client saves first, the stale write fails. Read the current integration and apply the change again. Each successful configuration update appends an immutable version.

Typed client

@marimo-hub/client uses the same generated OpenAPI document, so paths, parameters, bodies, and responses are checked against the live routes. apiData unwraps the envelope and throws an ApiRequestError on failure.

import { apiData, createApiClient } from '@marimo-hub/client';

const api = createApiClient({
	baseUrl: 'https://hub.example.com',
	headers: { Authorization: `Bearer ${process.env.MARIMOHUB_TOKEN}` },
});
const user = await apiData(api.GET('/api/v1/me'));

The exported types (Project, NotebookMeta, NotebookDetail, Session, Version, ResolvedUser, plus the full paths / components / operations) come straight from the schema. The marimohub SPA itself consumes this client.