Commands and MCP

The browser, CLI, HTTP API, and MCP adapters call the same workspace-scoped service. They share validation, permissions, revisions, review fingerprints, and recovery semantics. There is no arbitrary remote shell, SQL endpoint, or caller-selected provider URL.

Discover exact inputs

From the repository with locked dependencies installed:

npm run cli -- list
npm run cli -- schema project_get
npm run cli -- schema activity_add

Discovery works without a running dashboard. The schema command is authoritative for exact names, required fields, bounds, and enums. These guides explain workflows rather than duplicating evolving JSON schemas.

Task	Contract guide
Actionable attention and bounded operational previews	Overview
Projects and resource context	Projects
Reviewed project suggestions, grouping and priorities	Organize repositories
Move a project between workspaces	Transfers
Repository metadata and related resources	Repositories
Check and retain evidence for a repository's linked operational resources	Operational coverage, `repository_coverage`
Read retained operational coverage and progress without contacting providers	Saved operational coverage, `repository_coverage_get`
Discover repositories and reconcile reviewed enrollment changes	Fleet enrollment
Reviewed expectation presets, exceptions and bulk changes	Bulk expectations
Goal and progress reports	Activity
Saved repository coverage and latest collection results	GitHub coverage
Read-only GitHub collection	GitHub evidence
Local observations	Publishing
Invites, roles, and first-owner setup	Access
Hookrelay, Endpoint Monitor, and secret operations	Hooks, Monitoring, Secrets

Call one command

Prepare a JSON file matching the inspected schema. For example, project_get requires workspaceId and projectId. Use the real identifiers from your authorized workspace, not names or guessed IDs.

npm run cli -- call project_get --input request.json

Use --input - to read command metadata from stdin. Secret values are not ordinary command inputs. Secrets documents its dedicated private-input boundary.

Configure HQ_URL and one deliberate credential mode through protected environment configuration:

Unattended clients use an enrolled HQ_TOKEN and the paired Access ingress credentials HQ_ACCESS_CLIENT_ID and HQ_ACCESS_CLIENT_SECRET
A human session uses short-lived HQ_ACCESS_TOKEN instead; clear the automation variables before changing modes
Isolated local development uses an explicit loopback URL and --dev, which suppresses real credentials

The client rejects mixed modes, insecure remote transport, and credential-bearing redirects. The Access service pair admits a request at the edge; the HQ token still needs the exact workspace scope. See automation access.

MCP transports

The hosted MCP endpoint is /mcp on your protected application origin. A stdio adapter is available from the checkout. Invoke the executable directly so package-manager banners do not enter the protocol:

npx tsx cli/mcp.ts

For the isolated development server:

npx tsx cli/mcp.ts --url http://127.0.0.1:5178 --dev

Tool discovery exposes the shared bounded commands and their input schemas. A tool's presence is not permission to execute it. Reader, Reporter, and publisher credentials remain intentionally narrow. Hosted MCP cannot read your local files or environment; the trusted local Secrets input helper has a separately constrained interface.

Retries and reviews

Retain the original ID and exact content after an ambiguous submission. Do not generate a new review, report, or refresh ID just to retry the same operation. A changed input needs a new intentional operation. Consequential actions require a matching actor, authority, revisions, fingerprint, and unexpired review. Inspect existing receipts before requesting recovery or replay.