Conducted MCP

Conducted MCP

An MCP server that helps an AI agent plan and track a software project the way a disciplined team would — and it was built using the very methodology it ships.

Conducted MCP exposes Conducted Development — a lightweight, intent-driven methodology — to any MCP-capable agent (Claude Desktop, Cursor, and others). It is a stateless advisor + validator: it holds no project data and touches no files or git. The agent does all the I/O; the server supplies judgment — validate this artifact, is a standup due, what's the procedure for this phase, does this decision belong in the log.

The differentiator: this repository's own work/ folder — goal briefs, intent docs, a decision log, and cycle standups — was produced under the methodology the tool implements. It is the proof of process, not a sample.

What it does

A connecting agent gets, on demand:

• A guided kickoff (kickoff prompt + kickoff_questions / kickoff_plan tools) — a branching Q&A that bootstraps a project's planning structure for greenfield or existing codebases. For existing code the agent inspects the repo and the server pre-fills answers so the human confirms rather than authors from scratch.
• Strict artifact validation (validate_artifact) — submit a goal brief / intent doc / session log / standup, get back { valid, missing, warnings }.
• Phase procedures (next_procedure) — the ordered steps, what to read first, and the escalation points for wherever the agent is in the loop.
• Mechanical rule checks (standup_due, evaluate_gate, decision_log_guidance) — the rituals a solo practitioner most often lets slide, as stateless judgments over supplied facts.

The methodology's guides, templates, and conventions are served as read-only resources (conducted://guide/*, conducted://template/*, conducted://conventions) so an agent can learn the rules in-band.

Why it's built this way (Model C)

The server cannot enforce — an agent always has direct file access. So instead of pretending to be a gatekeeper, it is an advisor: pure functions returning judgments and procedures, no side effects, nothing to host with no data and no auth-to-data. That makes it portable, trivially testable, and cheap to run locally or remotely. The reasoning is written up in docs/DESIGN_SKETCH.md and the resolved trade-offs in DECISIONS.md.

Quick start

Heads-up: the npm package name is being finalized; until it is published, the npx invocation below is a placeholder. You can run it from a local clone today (see Development).

Add the server to your MCP client. Claude Desktop (claude_desktop_config.json) or Cursor (.cursor/mcp.json):

{
  "mcpServers": {
    "conducted": {
      "command": "npx",
      "args": ["-y", "conducted-mcp"]
    }
  }
}

Then ask your agent to "run the Conducted kickoff for this project," or call any tool directly.

Demo

A short recording of the kickoff flow bootstrapping a project will live here.

Development

npm install
npm run build     # bundles the methodology text, then strict tsc
npm test          # vitest
npm run lint      # eslint + prettier
npm start         # run the stdio server

The server is TypeScript on the official @modelcontextprotocol/sdk, ESM, strict mode. See CONTRIBUTING.md for the layout and conventions.

The methodology, in the repo

• work/goal-briefs/ — the goal briefs that drove this build
• work/intent-docs/ — one per ticket, the per-session contracts
• work/standups/ — cycle-gate standups
• work/decision-log.md — the append-only record of decisions
• AGENT_CONVENTIONS.md — how every session runs, model- and tool-agnostic

License

MIT © 2026 Jonathan Mostov