TruthGate for Agents

TruthGate is a fail-closed preflight, approval, evidence, and verification runtime for Hermes, OpenClaw, n8n agents, and custom LLM systems.

It does not promise zero hallucinations. It prevents unsupported output from being treated as verified completion.

v0.2 security model

• Server-owned policy cannot be weakened by request payloads.

• API routes require a bearer key.

• High-risk approvals are HMAC-signed, task-bound, expiring, and single-use.

• External tool traces can require HMAC signatures.

• Evidence IDs, citations, trust, claim coverage, tool success, output text, and artifact hashes receive deterministic checks.

• Evidence is scanned for common prompt-injection patterns before generation.

• A genuinely different verifier model is required by default.

• Audit events avoid raw prompts, tokens, and secrets.

• Verified memory is deduplicated and retains task provenance.

• Provider requests use bounded retries and sanitized errors.

Related MCP server: AI Workbench MCP

Install

python -m venv .venv
source .venv/bin/activate
pip install -e ".[dev]"
cp .env.example .env
truthgate-api

Open http://127.0.0.1:8787/docs. All /v1/* routes require:

Authorization: Bearer <TRUTHGATE_API_KEY>

Production startup rejects the included development secrets.

Hermes MCP integration

TruthGate exposes three narrow MCP tools:

• truthgate_preflight

• truthgate_verify

• truthgate_authorize

Use integrations/hermes/config.example.yaml to configure the stdio server in ~/.hermes/config.yaml, then run /reload-mcp in Hermes. Copy the companion integrations/hermes/SKILL.md into your Hermes skills directory if you want Hermes to follow the gate workflow consistently.

Human approval

Preflight returns task_id and task_digest. A human can issue a short-lived approval:

truthgate approve \
  --task-id TASK_ID \
  --task-digest TASK_DIGEST \
  --risk-level high

The token does not authorize a different task contract and cannot be consumed twice.

Deterministic rules

Requests can attach server-validated rules:

• evidence_exists

• tool_succeeded

• output_contains

• artifact_sha256

Model verification is never enough by itself. A final result is verified only when the independent verifier and all deterministic gates pass.

API

• GET /health

• POST /v1/preflight

• POST /v1/verify

• POST /v1/run

• POST /v1/authorize

• GET /v1/memory

• GET /v1/audit/{task_id}

Hard limitation

MCP cannot intercept every Hermes built-in tool result. Signed traces require a trusted host wrapper or a native Hermes tool-result hook. Without that hook, TruthGate should verify knowledge work and explicit evidence, but must not claim cryptographic proof of every action Hermes performs.