docs

Semel for coding agents

Connect a coding agent to Semel to create Build Briefs, run review cascades, and execute reviewed work — with least-privilege access and human approval where it matters.

View as Markdown
On this page

Semel gives your coding agent a governed way to turn intent into a reviewed, buildable specification. Connect Claude Code or Codex to https://mcp.semel.ai/mcp, authenticate with least privilege, and let the agent work through durable briefs, reviews, questions, and execution receipts while people retain the decisions only people may make.

What Semel is

Semel is an agentic product-development platform. Instead of a chat that forgets, Semel keeps a durable, versioned record of what you are building and why, and puts a review-and-approval workflow around the work an agent does. Your coding agent talks to Semel over the Model Context Protocol (MCP); Semel enforces who may do what, records every version and decision, and gates the moments a human must own.

When to use it

  • Turning an issue into a spec — paste Linear issue content into your request, or let a separately authorized harness fetch it, then draft a Build Brief you can inspect, edit, and version. Semel does not claim an embedded Linear connector.
  • Getting a spec reviewed — originate a hosted or local review and inspect its durable status, questions, and findings.
  • Governed local review — drive one claimed reviewer step under a human-approved, time-limited execution grant, using leases, fences, and ordered checkpoints.
  • Unattended automation — run a service agent with a scoped, revocable named token, with no secrets in chat or source control.

Brief → review cascade → execution

Three ideas carry most of the platform:

  1. Build Brief — the versioned source of truth for an effort. Every update replaces the complete canonical section set and mints an immutable version; it never edits history in place.
  2. Review cascade — an ordered set of reviewer lenses that reads a pinned brief version, records findings, and opens durable questions when a build-shaping decision belongs to a person.
  3. Execution protocol — a local reviewer claims one step under an explicit grant and advances it through ordered checkpoints, heartbeats, release, and submit. Epoch, fence, lease, and idempotency rules make restart and duplicate delivery safe; the final certificate attests what ran.

See Core concepts for the full vocabulary, and the Tool reference for the exact MCP tools your agent calls.

A five-minute path to first value

The fastest path is concrete and inspectable:

  1. Download https://docs.semel.ai/downloads/semel-install.mjs to a file, inspect it, and run the saved copy with Node.js 20+ and an explicit harness skill path.
  2. Register https://mcp.semel.ai/mcp in Claude Code or Codex and complete delegated OAuth.
  3. Start or reload the harness session, then require live serverInfo and a non-empty tools/list; a configured or OAuth label alone does not prove readiness.
  4. Ask the agent to list projects and briefs, then create a project and brief from pasted issue text. The write path requires brief:read plus brief:write and explicit confirmation of the resolved target.

Launch-supported harnesses today are Claude Code and Codex, each with a tested Quickstart. OpenClaw and standards-based generic MCP clients are documented as experimental until their install smoke tests exist — where a path is experimental, this documentation says so plainly.

What remains human

An agent may prepare work and request the next authorized action. It cannot approve or reject a gate, decide or commit/rebase/revert a proposal, mint or approve an execution grant, clear or destructively restart a review, change billing or spend, administer workspace credentials, connect Slack/GitHub/a tracker, consent to a model provider or BYOK terms, create a public share link, write to Brain, perform Done Check, or invoke the web-only /runs/:id/continue action. Linking an already-connected repository with project_codebase_link is the narrow codebase exception. Semel also never asks an agent to reveal chain-of-thought.