Supervise and orchestrate local coding agents (Muse/AGY) through a client-agnostic MCP tool surface with safe execution, Git worktree isolation, and deterministic verification.
List configured coding agents and their availability.
List configured repositories and safe metadata.
Start coding-agent tasks in a repository with agent, instruction, mode, and workspace strategy.
Continue an existing task session with follow-up instructions.
Get task state and concise metadata.
Read bounded task output with pagination cursor support.
Cancel running coding-agent tasks.
Get structured Git status for a task workspace or configured repository.
Get Git diffs for a task workspace or repository, including committed changes and untracked files.
Run configured verification profiles (e.g., test, lint) in a task workspace or repository.
Rely on SQLite-backed persistence, streaming logs, audit logs, and crash recovery.
Provides Git worktree isolation for coding-agent tasks and tools to inspect repository status and diffs, including committed changes since a task's base SHA and untracked files.
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type @ followed by the MCP server name and your instructions, e.g., "@coding-agent-mcp Start a task in my-repo to fix the failing tests and run verification."
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
Local, client-agnostic MCP server for supervising and orchestrating coding agents such as Muse and AGY.
The MCP server is the safe execution, isolation, and verification boundary. OpenAI/ChatGPT, Claude Desktop, custom MCP orchestrators, or any other compatible client act as external supervisors.
Safe Local Execution: Interacts with local agents without exposing arbitrary shell execution.
Git Worktree Isolation: Spawns tasks in isolated Git worktrees by default, protecting current work and enabling safe concurrency.
Client-Agnostic MCP Surface: Standard MCP tools compatible with any client:
list_agents
list_repositories
start_task
continue_task
get_task
get_task_output
cancel_task
get_repo_status
get_diff
run_verification
Deterministic Verification: Independent test and lint execution profiles returning real exit codes and outputs.
Agent Adapters: Pluggable adapters for Muse and AGY with headless execution, permission management, and session resumption. The optional agy-gemini adapter runs the installed AGY CLI through agy-gyro with an isolated Gemini API-key profile, avoiding account OAuth and Keychain.
Durable Persistence & Audit: SQLite-backed task state, streaming logs, and structured audit logs with crash recovery.
Related MCP server: Vibechemy
npm install
npm run buildCreate ~/.coding-agent-mcp/config.yaml (or copy examples/config.example.yaml):
server:
data_dir: ~/.coding-agent-mcp
max_concurrent_tasks: 2
repositories:
my-repo:
root: /path/to/my-repo
writable: true
default_workspace_strategy: worktree
verification_profiles:
test:
command: ["npm", "test"]
timeout_seconds: 900
lint:
command: ["npm", "run", "lint"]
timeout_seconds: 300node dist/index.js --config /path/to/config.yamlnpm testRuns the test suite covering unit policies, Git worktree isolation, task lifecycle, and end-to-end MCP tool invocations.
MIT
Cancels a running coding-agent task
Inputs: task_id (required).
Sends follow-up instructions to an existing task session
Inputs: task_id (required), instruction (required).
Returns the Git diff for a task workspace or configured repository (including committed changes since task base SHA and untracked files)
Inputs: staged, task_id, max_bytes, repository.
Returns structured Git status for a task workspace or configured repository
Inputs: task_id, repository.
Returns task state and concise metadata
Inputs: task_id (required).
Returns bounded task output with pagination cursor support
Inputs: task_id (required), cursor, max_bytes.
Returns configured coding agents and their local availability
Inputs: no named parameters.
Returns configured repository aliases and safe metadata
Inputs: no named parameters.
Runs a configured verification profile (e.g. test, lint) in a task workspace or repository
Inputs: profile (required), task_id, repository.
Starts a coding-agent task in a configured repository workspace
Inputs: agent (required), repository (required), instruction (required), mode, +1 more.
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Scored across 10 tools
Each tool addresses a distinct resource or action: agent discovery, repository discovery, task lifecycle management, output retrieval, git inspection, and verification. There is no meaningful overlap between get_task and get_task_output, or between get_repo_status and get_diff.
All tools follow a consistent snake_case verb_noun pattern (list_*, start_*, cancel_*, continue_*, get_*, run_*). The naming is predictable and makes the action-resource relationship clear across the entire set.
Ten tools is well-scoped for a coding-agent orchestration server. Each tool covers a necessary operation without redundancy, and the count is within the ideal range for maintainability and agent usability.
The tool surface covers the full task lifecycle: create (start_task), read (get_task, get_task_output), update (continue_task), and cancel (cancel_task). It also provides discovery (list_agents, list_repositories), repository inspection (get_repo_status, get_diff), and verification (run_verification), leaving no obvious dead ends for the stated purpose.
Control plane for autonomous software labor. Agents claim objectives over MCP with audit trail.
MCP server for your apps' tools and custom tools, plus hosted AI agents and approval-gated workflows
MCP server for AI agents to plan, verify, and deploy Cloudflare-native apps.
AI-native git hosting — repos, PRs, issues, CI gates, and AI code review over MCP (60 tools).
