Get the current picture.
Discover allowed repositories, inspect worktrees and setup state, read service status, follow jobs, and look at recent logs or supported configuration.
CANOPOD MCP / MODEL CONTEXT PROTOCOL
Your agent writes the code.
Give it the workspace, too.
Canopod MCP connects your coding agent to your local worktrees, services, and setup jobs. It can find out what’s running and help prepare a branch, without you relaying every detail.
15 tools in v0.5Local Streamable HTTPRead-only access availableOff until you enable it
WHAT YOU CAN ASK
Let the agent ask Canopod.
These examples show how a request maps to Canopod’s tools.
They’re illustrations of the workflow, not a live connection.
YOU ASK
“Which worktrees are running in the repositories I’ve shared with Canopod?”
canopod_repositoriesFind the repositories you’ve allowed.canopod_worktreesList their worktrees and branches.canopod_servicesRead the reported service states.
YOU ASK
“In my selected repository, create a worktree for feat/search from main and follow its configured setup.”
canopod_create_worktreeCreate the named branch’s worktree using repository defaults.canopod_jobWait for creation and any provisioning to finish.canopod_run_setupIf setup was skipped, run it and follow its job to completion.canopod_job_outputRead the setup output and report any failure.
YOU ASK
“Check the API service in my selected worktree. Show me its recent logs, then restart it.”
canopod_servicesFind the configured service and its state.canopod_service_logsRead recent, bounded service logs.canopod_restart_serviceRestart it, then follow the returned job.
“Selected” means identified in your conversation with the agent; MCP does not inherit the desktop selection. The agent needs the repository and worktree IDs. A running process still needs a readiness check.
THE TOOLS, IN PLAIN ENGLISH
A useful set of operations.
Access you can choose.
Prepare the next branch.
Create a worktree for a named branch and run your configured setup. Follow the job’s output instead of copying it from another terminal.
Manage what’s running.
Start, stop, or restart configured services in the selected worktree. Keep your agent and the desktop app working from the same backend.
Adjust supported settings.
Update supported repository fields or existing service settings with revision checks. This permission is separate from setup and service control.
CONNECT YOUR AGENT
A few choices in Settings.
Then back to your editor.
First, add your repository to Canopod. Then open Settings → MCP.
Choose the repositories.
Select the repositories the agent can access. Leave everything else outside its scope.
Choose its permissions.
Start read-only. Enable worktree setup, service control, or configuration when the task needs them, then select Enable MCP.
Connect your client.
Choose Claude Code or Codex under Connect to agent, connect it, and restart or reconnect the client.
Connect Claude Code from Canopod.
In Settings → MCP, select Claude Code and click Connect Claude Code. Restart or reconnect Claude Code, then use the first task Canopod provides to check the connection.
Use a current client. If you use a custom CLAUDE_CONFIG_DIR, follow the manual connection instructions.
Connect Codex from Canopod.
In Settings → MCP, select Codex and click Connect Codex. Restart or reconnect Codex so it loads the connection, then try the first task shown in Canopod.
Automatic setup preserves the client’s other settings. Use a current Codex installation.
Use a compatible local MCP client.
Open the manual setup instructions in Settings → MCP. Use the local Streamable HTTP endpoint and bearer authorization Canopod shows there. Keep that token private.
Remote exposure is unsupported. Client-specific setup varies; use the instructions for your MCP client.
A CLEAR SCOPE
You choose what
the agent can reach.
MCP is off by default and only listens locally. Repositories and capabilities are explicit choices in Settings.
How Canopod handles accessConfigured operations
Setup and service commands run from your configuration. There is no general-purpose shell tool. Review the commands and grants you enable.
Deliberate limits in v0.5
Dedicated worktree-removal and database-management tools aren’t exposed through MCP. Configuration patches cover supported fields, not adding or removing services.
Local connection, your client’s data policy
Canopod sends no telemetry. Your coding agent may send workspace context to its provider according to that client’s settings.
A few details worth knowing.
Is MCP the same as launching an agent in a terminal?
They are separate features. A terminal agent gets a task handoff and runs in a worktree. An MCP client calls the tools Canopod exposes, using the repositories and permissions you have allowed. You can use either workflow.
Can I run the MCP backend without the desktop app?
Yes. The packaged canopod-backend serve runs headlessly under your own process supervisor. In v0.5, it cannot own the same data directories at the same time as the desktop app. Read the headless setup guide.
Does my agent get the full log history?
No. Service logs are bounded and read tools can use cached state. Use the reported state and recent output as evidence, and perform a readiness check before treating a service as ready.
Where can I see all of the tool names?
The MCP tool reference lists the 15 tools and their required permissions. You can also inspect the implementation in the open-source repository.
Give your agent a better starting point.
Install Canopod, add a repository, and connect your agent when you’re ready.