Siphrix — see what your AI agents actually did
Your coding agent edits files, runs commands and reaches the network on
your behalf, all day, at machine speed. Siphrix writes every one of
those actions to a tamper-evident record and raises the risky ones as
warnings — without ever getting in the way.
Audit-first, by design: Siphrix does not block anything. It
observes, records and warns. (A frozen enforce mode exists behind an
explicit SIPHRIX_MODE=enforce opt-in; it is off, and the default
install never stops an action.)
An observer must not break the thing it observes.
One-click setup
Install the extension and run “Siphrix: Set Up” (also
offered automatically on first activation). One click:
- finds a Python 3.11+ interpreter,
- installs the Siphrix engine (from the wheel bundled in the
.vsix,
or from PyPI),
- runs
siphrix agent-setup — configures every detected agent:
the PreToolUse hook for Claude Code, advisory MCP tools for
Codex (~/.codex/config.toml),
- offers to open the live console.
No manual pip install, no editing settings.json, no environment
variables. From then on every agent tool call is recorded — what ran,
in which file, in which app — and risky ones surface as warnings.
What you get inside VS Code
Four commands. No server to start, no port, no token.
- Siphrix: Set Up — finds Python, installs the engine, wires Claude
Code and Codex. One click.
- Siphrix: Connect to My Account — paste a connection code from your
console and this machine reports into your account. It installs the
engine first if it is missing.
- Siphrix: What Did My AI Do? — the recent record in plain words:
who acted, where, on which file or command, and what was flagged.
Read straight from disk, so it works whether or not anything is
running.
- Siphrix: Open Console — your account, or the local console.
A status bar item says which of three states you are in: not set up,
recording locally, or recording and syncing.
Watch rules live in the console, where they apply to your whole account
rather than to one machine.
Works with your team
Siphrix runs entirely on your machine. Optionally, connect it to a
console (self-hosted, or siphrix.com) and the
record from every machine lands in one place: who acted, in which app,
on which file — plus warnings, weekly reports, and an integrity check
that proves the record was never altered.
siphrix cloud-connect --server https://siphrix.com --code <your code>
Rules flow console → machine. A machine can report what it watches, but
can never write policy for the organization: visibility flows up,
authority flows down.
Settings
One setting, and you will probably never touch it.
| Setting |
Default |
Notes |
siphrix.pythonPath |
auto-detected |
The interpreter used for setup. Set it only if your Python is somewhere unusual. |
Privacy & posture
- Reads only your own record, from
~/.siphrix on this machine. No
server, no port, no network call of its own.
- Never executes an action, and never blocks one by default.
- Records metadata only — action type, target path, command line,
verdict, time. Never file contents, never message text. Anything
secret-shaped (
token=…, api_key=…) is masked before it is stored
or synced.
- No telemetry. Nothing leaves your machine unless you explicitly
connect it to a console you chose.
Build from source
cd tools/vscode_extension
npm install && npm run compile
npx @vscode/vsce package # -> siphrix-vscode-<version>.vsix
code --install-extension siphrix-vscode-<version>.vsix
Source-available under the Business Source License 1.1 from 2.0.0 of
the Licensed Work (see the LICENSE file in this directory). Versions of
this extension published before the pivot were released under MIT and
remain MIT permanently.