Practical use.
MCP is the in-session path: the connected coding tool can request a focused ShardStitch operation while the session is still available. The local CLI or app remains useful when no MCP client is connected.
49 MCP tools.
Claude Code, Cursor, Devin, and Windsurf can call ShardStitch directly through MCP tools for scan, recall, handoff, council, and recovery workflows.
What it carries.
- Scan and recall tools
- Recovery packet generation
- Council and verification workflows
- Local project memory access
How it fits ShardStitch.
This feature exposes recovery actions through MCP so a supported coding tool can request project context without relying on a pasted transcript.
When MCP is the right path.
Use MCP while a supported AI session is still alive and can call local recovery actions directly. A scan can prepare evidence before a context limit, while recall and handoff tools can keep the next step close to the repository. If the session is already dead, use the local dashboard or CLI instead.
MCP availability varies by tool and configuration. Inspect the generated server settings, keep the server local where appropriate, and verify each result before treating it as project truth.
Use MCP before the session becomes unreliable.
For an active supported client, ask it to scan or prepare a handoff at a meaningful checkpoint; review the returned files and evidence before the agent continues.
- Confirm the MCP server is connected in the intended project and that the requested tool is available to the client.
- Request the narrow operation needed, such as scanning current project state or creating a continuation packet.
- Inspect the response against the checkout, then save or inject the handoff only into the intended project pickup file.
Worked example.
A useful checkpoint is before a context limit or tool switch: ask the connected client to scan the active repository and create a focused handoff, then inspect the result before the current session ends. If the client is disconnected or the project path is wrong, use the local app or CLI and verify the same repository. Keep the generated file in the intended workspace.
Limits and verification.
MCP availability depends on client configuration and a running local server; it is not a promise that every AI target can call all tools. A successful MCP response does not prove that later edits are correct. Verify the resulting diff and tests as usual.