Get to a verified result quickly
EVALUATE
Use synthetic or authorized text in the playground. Inspect the protected context, route decision, and receipt without an external model call.
SELF-HOST
Run the gateway beside the source data, configure a policy, and protect one request inside your own boundary.
CONNECT AN APP
Point a non-streaming OpenAI client at the gateway and bind each call to a narrow task, policy, and privacy mode.
CONNECT AN AGENT
Install five allowlisted tools plus the fail-closed Agent Skill for OpenClaw, Hermes, Codex, Claude Code, or another compatible host.
One privacy boundary, several ways to use it
- 01
DETECT
Find direct identifiers, exact values, wallets, and credential-like material.
- 02
MINIMIZE
Use the declared task to remove facts the downstream job does not need.
- 03
TRANSFORM
Pseudonymize, generalize, conceptualize, remove, block, or keep local according to policy.
- 04
ROUTE
Send only approved protected context—or stop when the correct route is local_only.
- 05
GUARD
Inspect model output before any authorized application-side reconstruction.
- 06
RECEIPT
Record policy, counts, route, timing, reduction, and integrity without storing source content.
Know what is live before you integrate
| SURFACE | STATE | USE IT FOR |
|---|---|---|
| Browser playground | LIVE | Deterministic product evaluation with local browser processing |
| Gateway + Vault | SELF-HOSTED | Protected API and OpenAI-compatible application workflows |
| MCP server | SELF-HOSTED | Five agent-safe tools with reconstruction withheld |
| Agent Skill | LIVE | Portable fail-closed operating rules for compatible agents |
| Virtuals ACP | HOLD | Blueprint only; identity, signer, settlement, and recovery are unverified |
| x402 + $SHRED payment | HOLD | No posted rate, token contract, recipient, or settlement receipt |
Docs for humans and runtimes
The Build section provides copyable setup paths for generic MCP hosts, OpenClaw, Hermes Agent, OpenAI Codex, Claude Code, and OpenAI-compatible clients.
Use the portable Agent Skill to preserve purpose binding, fail-closed behavior, the no-reconstruction tool boundary, and honest receipt and pricing reporting across runtimes.