I build the layer that lets LLMs act on real systems without acting unsupervised, with an audit trail for every action.
Approve opens a discussion thread on this repo. Deny goes to my repositories.
One tool registry, three doors (chat, external agents, approvals) and one ledger that records every call. The approval gate in the diagram is the card above.
flowchart LR
U[User] -->|SSE stream| A[Chat agent]
X[Codex or Cursor] -->|MCP with OAuth2| M[MCP server]
A --> R[Tool registry]
M --> R
R --> Q[Read tools]
R --> G{Approval gate}
G -->|approved| W[Write tools]
Q -.-> L[(Audit ledger)]
G -.-> L
W -.-> L
D[Your docs] --> I[Index with citations]
I --> Q
classDef gate stroke:#9a6700,stroke-width:3px
class G gate
| Block | The rule |
|---|---|
| Tool registry | One definition per tool. Chat and MCP both read from it, so they never drift. |
| MCP server | OAuth2 per user. Never a shared key. |
| Approval gate | Reads run freely. Every write waits for a person, and the wait survives a resume. |
| Audit ledger | Every tool call, approval and denial is recorded once and never edited, with who, what and when. Compliance reads the ledger, not chat logs. |
| Index | An answer points at a file and a passage. A wrong citation fails a test. |
- Tool errors go back to the model as structured results it can recover from, never as stack traces.
- Mask sensitive values before they reach the model provider, not after.
- Stream the answer as it is produced. Buffering hides every failure until the end.
Your system has a different shape? Approve the card above and tell me which block breaks.





