keshro-mcp 0.1.38

High-stakes engineering project execution for AI agents — MCP server

Keshro MCP

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

Tool	What it does
preview_plan	Run Keshro's pre-plan intake and clarifying-question preview
generate_plan	Generate a plan from a description using AI
list_plans	List all plans
get_plan	Get a plan with all tasks
plan_status	Progress summary (task counts, enrichment sources)
next_task	Get the next actionable task
create_plan	Create a plan manually
start_task	Mark a task as in progress
complete_task	Mark a task as done
block_task	Mark a task as blocked
unblock_task	Clear a blocker
append_task_note	Add a note to a task
add_task_artifact	Attach an artifact link
record_decision	Log a decision with context, choice, and reasoning
edit_task	Edit task title or description
push_to_tracker	Push tasks to Linear, Jira, or GitHub as issues
sync_pull	Pull status updates from connected issue tracker
export_project	Export project data

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.