Connect
MCP and external AI tools
Connect an external local AI client and control what it may do to instances.
Two versions of this page follow. Step by step walks through the task in order. In depth explains the underlying behavior. They describe the same product, at different levels of detail.
Step by step
Follow the steps in order. No prior modding knowledge assumed.
MCP is an advanced optional bridge that lets a compatible AI tool inspect Agora and request supported actions. Most users do not need it; the integrated AI Assistant is simpler for occasional help.
After this you will be able to
- Decide whether MCP is appropriate for your workflow.
- Start and stop the local server.
- Understand the security impact before connecting a client.
Choose integrated AI or MCP
Use Integrated AI for GitHub Copilot chat inside Agora. Use MCP when you already operate an external AI client and want it to inspect instances, read bounded crash evidence, search local knowledge, or request supported enable/disable actions through Agora's tool interface.
Connect a client
Enable Advanced mode and AI / MCP Server in Settings. The desktop server binds only to 127.0.0.1:39741 and exposes SSE at /sse plus streamable HTTP at /mcp. Every request must send the generated token in the Authorization: Bearer <token> header; query-string tokens are not accepted.
- Start the MCP server only when needed.
- Copy the configuration for Kilo Code, Opencode, Claude Desktop, or another compatible client.
- Store the Bearer token like a local credential.
- Confirm the client can list instances before allowing any modifying action.
- Stop the server when the workflow is complete.
Approve changes carefully
An AI request is not proof that an action is safe. Review the target instance and requested operation. Keep important instances locked and use snapshots before permitting changes.
In depth
How the system actually behaves, and why, for readers who prefer the model to a recipe.
Operate MCP as a privileged local automation surface: authenticate clients, scope approvals per instance, monitor state, and preserve transactional recovery.
After this you will be able to
- Configure client authentication and rotate credentials.
- Apply least privilege to modifying tools.
- Troubleshoot connectivity without weakening the local boundary.
Secure the connection
Use http://127.0.0.1:39741/sse for SSE or http://127.0.0.1:39741/mcp for streamable HTTP. Never expose either endpoint through port forwarding or a public reverse proxy. Send the token only in the Authorization header. Regeneration invalidates the previous token and requires updating every client.
Scope tool approvals
The ten tools are list_instances, list_instance_mods, read_mod_manifest, get_system_context, search_crash_signatures, search_knowledge_base, disable_mod, enable_mod, read_latest_crash, and suggest_mod_incompatibility. Use per-instance approval settings for modifying tools. Read-only inspection is lower risk but can still expose instance names, file metadata, or crash context to the connected AI provider.
Troubleshoot methodically
Check that the server status is running, the URL uses 127.0.0.1 and port 39741, the client transport is SSE, the token is current, and local security software is not blocking the process. Do not disable system security or bind the server broadly to solve a client configuration error.
This guide is the same text Agora shows in its built-in Help & Guide, which is searchable and links directly to the screens it describes.