The route of a tool call

The animation explains the architecture; no live actions are executed.

A result carries its source ID, time and coverage.

AI client

• User question + model

The AI client turns the question into an available tool call.

MCP interface

• tools/list · tools/call + schema

MCP describes the function, input and any structured output.

OpenClaw

• Routing + workflows

The gateway organises execution behind the MCP connection.

Adapters with bounded contracts

• Source ID · time · coverage · traceable context

An adapter retrieves a bounded result and identifies the source it read.

Local app data

• SQLite · caches · exports
• Read-only where appropriate

Local app data needs a reading route that accounts for caches and source freshness.

Native automation

• AppleScript / JXA / frameworks
• Actions through suitable app interfaces

Native automation uses available app functions and macOS permissions.

My knowledge base

• PostgreSQL + pgvector + graph
• Sources and relationships stay separate

The knowledge base exposes indexed sources and recorded relationships.

Retrieve → combine → rerank → select evidence → answer the question

The search layer combines candidates and selects evidence for the answer.

Maintenance loop

• Monitor interface → scope change → Codex → tests → review → apply → verify

When something changes, I check the adapter, contract and original question.

Connections

• AI client → MCP interface
• MCP interface → OpenClaw
• OpenClaw → Adapters with bounded contracts
• Adapters with bounded contracts → Local app data
• Adapters with bounded contracts → Native automation
• Adapters with bounded contracts → My knowledge base
• Local app data → Retrieve → combine → rerank → select evidence → answer the question
• My knowledge base → Retrieve → combine → rerank → select evidence → answer the question
• Retrieve → combine → rerank → select evidence → answer the question → Maintenance loop
• Maintenance loop → Recheck → Adapters with bounded contracts

How to read this

1. The AI client turns the question into an available tool call.
2. MCP describes the function, input and any structured output.
3. The gateway organises execution behind the MCP connection.
4. An adapter retrieves a bounded result and identifies the source it read.
5. Local app data needs a reading route that accounts for caches and source freshness. Native automation uses available app functions and macOS permissions. The knowledge base exposes indexed sources and recorded relationships.
6. The search layer combines candidates and selects evidence for the answer.
7. When something changes, I check the adapter, contract and original question.