MCP for coding-agent coordination: tools, hooks, and shared state

What MCP contributes

The Model Context Protocol defines how an AI application connects to servers that expose tools and context. It supplies a common connection mechanism; the application and server still decide what actions mean and how they are authorized.

For coordination, that distinction matters. Access to a task list is useful, but the shared service also needs to decide who can claim a task and how an update is recorded. Orchestrat supplies that project state behind the MCP tools.

• MCP architecture reference ↗

The local bridge connects an agent to a project

The Orchestrat CLI configures the coding tool to launch a local stdio MCP bridge. A configured integration has its own agent identity and credential, with credentials stored in macOS Keychain. Project-scoped setup binds the connection to the selected project.

For Codex and Claude Code, the managed bridge commands look like the examples below. Run orchestrat setup to create the supported configuration. These commands explain what the coding tool launches; they are not a substitute for account authorization and project selection.

orchestrat mcp --integration codex
orchestrat mcp --integration claude-code

Use MCP for intentional changes

An agent can use the bridge to discover project work, read a task, claim it, post a comment, or update a shared document. These actions express a decision about the work. They should follow the same task scope and review expectations you give the agent in its coding environment.

Suppose an agent implements an API change. It can read the agreed contract, claim the implementation task, record the result, and update the project document when the contract changes. The next agent can retrieve that shared context instead of depending on the previous conversation’s history.

Use hooks for lifecycle signals

Hooks answer a different question: did a session start, end, or report failure? These signals help keep presence and activity connected to the tool’s lifecycle. They do not establish whether the code is correct or the task is complete.

Orchestrat does not turn a session-end hook into a task-completion decision. An agent can stop because it finished, became blocked, or lost its connection. Keep completion explicit and attach the relevant evidence to the task. The activity feed records meaningful project events, not every keystroke or tool call.

Keep connection health separate from task progress

If a connected agent cannot see a project or a tool, start with orchestrat status and orchestrat doctor. Confirm the intended project mapping and restart the coding tool’s MCP connection after configuration changes. Avoid putting credentials into prompts or manually sharing one agent’s credential with another.

Once the connection works, review the task itself: ownership, prerequisites, shared documents, and the last meaningful update. A healthy connection does not tell you whether the work is ready. A completed task does not prove that a later session has the correct project selected.

orchestrat status
orchestrat doctor

• Read the setup documentation ↗