TechLead MCP

@gpriday/techlead-mcp is a read-only MCP server that gives coding agents two structured tools:

• techlead.plan

• techlead.review

It supports local stdio mode, uses provider-side structured outputs, and can route across OpenAI, Anthropic Claude, and Gemini.

Each tool response includes the structured plan/review, provider token usage when returned by the API, and an estimated USD provider cost derived from the configured per-model pricing catalog. Provider billing dashboards remain authoritative.

Install

npm install -g @gpriday/techlead-mcp

Related MCP server: greybeard

Configure

Create a .env with at least one provider key:

OPENAI_API_KEY=
ANTHROPIC_API_KEY=
GEMINI_API_KEY=

Optional config:

techlead-mcp init

Optional GitHub issue task source:

TECHLEAD_GITHUB_TOKEN=github_pat_or_fine_grained_token

Then callers may omit task and provide:

{
  "cwd": "/repo",
  "githubIssue": {
    "repository": "owner/repo",
    "issueNumber": 123
  },
  "files": []
}

The server reads the issue body and issue comments through the GitHub REST API and uses that text as the task.

Optional cheap-model router:

TECHLEAD_OPTIONAL_LLM_ROUTER=true
TECHLEAD_ROUTER_PROVIDER=openai
TECHLEAD_DEFAULT_MODEL_TIER=balanced

When enabled, the router model receives the resolved task text, including GitHub issue text when githubIssue is used, plus context size and risk signals. It may choose OpenAI, Anthropic, or Gemini for the actual plan/review call, but deterministic safety rules still force max-tier routing for auth, security, migrations, data loss, failing tests, and similar high-risk work.

Run

techlead-mcp serve
techlead-mcp models

TechLead MCP is stdio-only because it is designed to run next to the repository and read safe files under configured local roots. By default, local reads are restricted to the process working directory. Do not run the server from a sensitive parent directory.