# claude-tokenization-proxy

A local Data Loss Prevention (DLP) proxy that sits between `claude-code` and the
Anthropic API. It scans outgoing requests for secrets with
[GitGuardian's `ggshield`](https://github.com/GitGuardian/ggshield), swaps each
secret for an opaque placeholder token before it leaves your machine, then
restores the real secret in the streamed response on the fly — so Claude never
sees your secrets but tooling that echoes a placeholder still gets the real
value back.

```
claude-code ──POST /v1/messages──► proxy ──redacted──► api.anthropic.com
                                     │  (ggshield scan, secret → <<<gg_token_…>>>)
claude-code ◄──SSE (detokenized)─── proxy ◄──SSE stream──┘
                                     (<<<gg_token_…>>> → real secret, streaming)
```

## How it works

1. **Request interception** — the JSON body of `POST /v1/messages` is scanned by
   `ggshield`. Every detected secret is replaced with a unique
   `<<<gg_token_[hex]>>>` placeholder, and the `token → secret` mapping is kept
   in memory (`src/store.ts`). The redacted body is forwarded upstream.
2. **Streaming detokenization** — the Anthropic SSE response is piped through a
   stateful `Transform` (`src/detokenize.ts`) that scans the byte stream for
   placeholders and substitutes the original secret back. It holds back only the
   few characters that could still become a token, so a placeholder split across
   chunk boundaries (`…<<<gg_to` / `ken_abc>>>…`) is reassembled correctly.

Secrets live only in memory for the lifetime of the process — restart the proxy
and the mapping is gone.

## Requirements

- **Node.js ≥ 23.6** (the proxy runs TypeScript directly via Node's built-in
  type stripping — no build step, no `ts-node`/`tsx`). On Node 22.6–23.5 run
  with `node --experimental-strip-types src/index.ts`.
- **`ggshield`** on your `PATH` and authenticated (`ggshield auth login`). If
  `ggshield` is missing or errors, the proxy logs a warning and forwards traffic
  **unredacted** (fail-open) — it never blocks your requests.

## Setup

```bash
npm install        # dev-only deps (TypeScript types); not needed to run
npm start          # starts the proxy on http://127.0.0.1:8080
```

Point `claude-code` at the proxy:

```bash
export ANTHROPIC_BASE_URL=http://127.0.0.1:8080
claude
```

For the [fish shell](https://fishshell.com/):

```fish
set -gx ANTHROPIC_BASE_URL http://localhost:8080
claude
```

That's the only change `claude-code` needs — it keeps using your normal
`ANTHROPIC_API_KEY` / `x-api-key`, which the proxy forwards untouched.

## Configuration

| Env var                 | Default            | Purpose                          |
| ----------------------- | ------------------ | -------------------------------- |
| `PORT`                  | `8080`             | Port the proxy listens on        |
| `ANTHROPIC_UPSTREAM_HOST` | `api.anthropic.com` | Upstream API host to forward to |

## Tell Claude to leave the tokens alone

So the model preserves placeholders verbatim (e.g. when writing them to files),
copy [`templates/CLAUDE.md`](templates/CLAUDE.md) into the `CLAUDE.md` of any
project you run `claude-code` in:

```markdown
# Security Constraints

Do not modify, remove, or evaluate strings formatted as `<<<gg_token_[a-z0-9]+>>>`.
Treat them as opaque identifiers and preserve them exactly as they appear when
writing or reading files.
```

## Scripts

| Command           | What it does                                          |
| ----------------- | ----------------------------------------------------- |
| `npm start`       | Run the proxy                                         |
| `npm run dev`     | Run with `--watch` (restarts on file changes)         |
| `npm test`        | Run the detokenizer state-machine self-check          |
| `npm run typecheck` | Type-check the source with `tsc --noEmit`           |

## Limitations

- Detection is only as good as `ggshield`. It catches known secret patterns and
  generic high-entropy strings, not arbitrary sensitive prose.
- Re-inserted secrets are assumed to be escape-neutral (typical API
  keys/tokens). A secret literally containing a backslash or double-quote could
  mis-escape inside the response JSON. See the `ponytail:` note in
  `src/detokenize.ts`.
- The token map is in-memory only; it is not shared across proxy restarts or
  multiple proxy processes.

## License

Copyright (C) 2026 Rootiest. Licensed under the
[GNU Affero General Public License v3.0 or later](LICENSE).