MCP host integration
Configure a stdio host and understand the four model tools and fixed identity scope.
Configure the host
Point the MCP command to the built rsia executable. Adapt the outer configuration to the host:
{
"mcpServers": {
"rsiagent": {
"command": "/absolute/path/to/rsia",
"args": ["mcp", "--data", "/absolute/path/to/rsia.sqlite3", "--namespace", "personal", "--actor", "my-assistant"]
}
}
}{
"mcpServers": {
"rsiagent": {
"command": "/absolute/path/to/rsia",
"args": ["mcp", "--data", "/absolute/path/to/rsia.sqlite3", "--namespace", "personal", "--actor", "my-assistant"]
}
}
}Reconnect and confirm the four tools. A successful connection does not automatically wire the host's task context and execution receipts into the complete improvement cycle.
Model tools
| Tool | Purpose | Boundary |
|---|---|---|
evo_prepare | Return a fixed snapshot of approved active skills | Does not execute the host task |
evo_feedback | Record reported run feedback | Self-report is not independent scoring |
evo_propose | Submit a bounded proposal for validation | Does not approve or publish it |
evo_inspect | Read an authorized redacted object | Does not grant admin or raw-data access |
Use the schemas advertised by the current server. Do not invent identity fields to elevate a request.
Identity and verification
The stdio entry fixes the Agent identity at startup. Namespace scopes data and actor identifies the caller. Management requires a separate entry and credential; see management.
Verify initialization and tools/list, prepare a snapshot, confirm that the host actually attaches it, execute the task and send feedback. See the lifecycle for stronger evidence than successful tool calls.