The intelligent execution layer for coding agents, exposed as an MCP server for high-stakes engineering projects.

pip install keshro-mcp

When to use this vs the CLI

Use the CLI (pip install keshro) for the full experience: interactive clarifying questions, migration detection, parallel execution in isolated worktrees, git checkpoints, cross-task context routing, and cost tracking.

Use MCP if your agent platform speaks MCP and you want direct tool-call access to Keshro plans and tasks.

The CLI gives you more control. MCP is more flexible for custom integrations.

Setup

Set your API token:

export KESHRO_API_TOKEN="ksh_pat_..."

Get one from keshro.com/account.

Connect to your agent

MCP works with any agent that supports the protocol — Claude Code, Cline, Continue, Zed, and others.

Claude Code — add to ~/.claude.json:

{
  "mcpServers": {
    "keshro": {
      "command": "keshro-mcp",
      "env": { "KESHRO_API_TOKEN": "ksh_pat_..." }
    }
  }
}

Other MCP clients — point your client at the keshro-mcp binary with KESHRO_API_TOKEN set in the environment. The server uses stdio transport.

Available tools

Current parity notes

MCP now supports the newer task controls exposed in the web product:

• explicit depends_on task dependencies
• parallelizable task scheduling hints
• per-task executor selection
• generic issue linking via issue_id, plus external issue fields
• pre-plan intake via preview_plan

It still remains thinner than the CLI for actual execution orchestration. The CLI owns parallel local worktrees, git checkpoints, richer execution transcripts, and the direct keshro continue runtime loop.

License

MIT

Releases

Publish the MCP package with one GitHub Actions run after you bump pyproject.toml:

gh workflow run "Publish MCP"

That workflow reads the package version from pyproject.toml, publishes the package to PyPI, then creates the matching vX.Y.Z GitHub release automatically.