MCP

MCP is how Eyeban grows beyond built-in tools. Wire servers in JSON. The agent discovers their tools with GetMcpTools and calls them with CallDynamicTool, under the same Ask / Allow loop as Shell.

Files

• User: ~/.eyeban/mcp.json
• Project: .eyeban/mcp.json

Later paths override the same server name, so project wins over user. You can also add servers in Settings; they save to the user file.

stdio

A stdio server is a local command:

{ "mcpServers": { "example": { "command": "npx", "args": ["-y", "@modelcontextprotocol/server-filesystem", "/path"] } } }

Optional env injects environment variables. Do not commit tokens.

SSE / HTTP

A remote server uses url. If url is set, transport defaults to sse unless you set transport explicitly. Settings exposes this as “SSE / HTTP”.

{ "mcpServers": { "tracker": { "url": "https://mcp.internal.example/sse", "transport": "sse" } } }

Resources

When a server exposes resources, the agent can read them (ReadMcpResource). Prefer servers you can audit.

Trust

There is no curated Eyeban store and no unsigned “plugin pack.” Read the tool schema before Allow all. See Extensions for hooks and the trust boundary.