● source-available fair-code · Core beta

Get started

Promtect Core beta is installable today under the Sustainable Use License. Install it, then use the offline proof commands below to see exactly what it masks.

1. Install

brew install Amitk3293/tap/promtect
promtect --version
promtect selftest

Prefer to build it yourself? Clone the repository and run cargo install --path . from the clone. Binaries are published as attested GitHub Releases; verify the SHA-256 sidecar before installing.

2. Run your tool through it

guard ollama is the runtime-proven path. Claude and Codex guard integrations are beta and must not be treated as launch-ready until their integration gates pass.

promtect guard ollama run deepseek-r1

Prefer manual? Start promtect and set the tool's base URL to http://127.0.0.1:8790 (or …/v1 for OpenAI-style tools). See the integration guides.

3. What happens to a request

Mask outbound. Promtect replaces recognized detector matches in supported, uncompressed UTF-8 request bodies before forwarding.

Forward masked. The remaining prompt goes to your configured upstream. Authorization and applicable end-to-end headers are forwarded without body scanning; Host, Content-Length, Transfer-Encoding, and Accept-Encoding are removed or regenerated.

Restore inbound. An unchanged sentinel can become the original value as the reply streams back, or remain masked in strict mode (PROMTECT_RESTORE=false).

Prove it's working (offline)

promtect selftest                 # masks a canary, proves no leak, restores it
printf '%s\n' 'ship it with AKIAIOSFODNN7EXAMPLE' | promtect mask
promtect playground               # full proxy round-trip vs a mock upstream, narrated

promtect playground sends a request carrying fake secrets through the real proxy and prints what the upstream actually received (masked) versus what you got back (restored). Nothing leaves your machine.

Configuration

Variable	Default	Purpose
`PROMTECT_PORT`	`8790`	Port Promtect listens on (loopback).
`PROMTECT_MODE`	`anthropic`	Upstream preset: anthropic, openai, ollama, openrouter.
`PROMTECT_UPSTREAM`	none	Explicit upstream URL; overrides mode (chaining).
`PROMTECT_RESTORE`	`true`	`false` = strict mode (never re-insert secrets).
`PROMTECT_BLOCK_RISKY`	`false`	`true` = refuse high-risk/unverified upstreams.
`PROMTECT_AUDIT`	`promtect-audit.jsonl`	Beta audit path; records are designed to omit detected values.

Full documentation

FAQ & troubleshooting →

Setup gotchas, multi-tool, strict mode, "is it actually masking?"

Detector reference →

Every detector, by provider.

Architecture →

How detect → vault → mask → restore fits together.

Integrations →

Claude Code, Cursor, Codex, Ollama, OpenRouter, chaining.

Threat model →

What it protects, and what it deliberately doesn't.

Roadmap →

What is done, what is next.