Documentation
Surfil is an on-device control plane that runs beside the AI coding agents you already use - Claude Code, Cursor, Codex, Copilot, and MCP-based agents. One install intercepts at a single point, runs every request through a five-stage pipeline, and writes what it learns to one signed knowledge network, Weave.
The shape of the system
Everything sensitive happens on your machine. The on-device Rust runtime intercepts each agent request at one point - your agent's base URL - and runs it through five stages, in order: security → cost → quality → observability → memory. The edge (stateless workers at api.surfil.com) only authenticates, meters, and signs. Your source code never crosses that boundary.
agents ─→ interceptor ─→ security → cost → quality → observability → memory ─→ Weave
(your device, one point) (signed facts)How these docs are organized
- Getting started - Quickstart (install to first verified receipt) and Install (per-agent notes, what changes on disk, byte-exact restore).
- Concepts - the four mechanics every product builds on: zero-trace, Weave, signed outputs, and Credits.
- Guides - task-shaped walkthroughs: cut costs, block risky calls, use local memory, verify a receipt offline.
- Reference - the public edge API and the FAQ.
Where to start
If you want Surfil running in the next few minutes, go straight to the Quickstart. If you want to understand what you are installing before you install it, read Zero-trace first - it is the invariant the rest of the system is built around.
What these docs promise
- Every claim traces back to something you can measure or verify yourself. Nothing is "certified"; savings are reported as a measured floor, in tokens.
- Every command shown is a real command. Where an exact flag or config path is still being finalized, the page says so instead of guessing.
- Everything Surfil changes on disk is reversible: uninstall restores your original agent configs byte-for-byte.