Pre-action safety gate for AI agent tool calls.

Call it before your agent executes. It returns a verdict — COMMIT or NO_COMMIT — with a confidence score, a reason, and a tamper-evident hash chain entry, so you have a defensible record of why an action was allowed or stopped.

Blocks unsafe money movement Blocks unreviewed code execution Blocks PII leakage Pay-per-call via x402

See endpoints & pricing → OpenAPI spec Swagger UI

verdict log — example

14:02:11Z /evaluate/fast COMMIT transfer within agent's approved daily limit

14:02:47Z /evaluate/safety NO_COMMIT destination not in allow-listed wallet set

14:03:02Z /evaluate/jailbreak NO_COMMIT forbidden pattern matched: role-override attempt

Illustrative example of the response shape — not production data. Every verdict is written to an append-only, hash-chained log; anyone holding the record can verify it wasn't altered after the fact.

What it's for

Financial agents

Stop a transaction before it settles, not after.

Tool-using agents

Gate shell/API calls the agent is about to make.

Content & support agents

Catch jailbreaks and PII leakage before output ships.

Endpoints

/evaluate/fast Low-latency single-shot check — agent_id + response text in, verdict out. tool-call-safety

/evaluate/strict Full policy audit against your configured ruleset, higher confidence. policy-audit

/evaluate/jailbreak Detects prompt-injection and role-override attempts in agent context. jailbreak-detection

/evaluate/safety Checks a proposed action (transfer, call, write) against safety constraints. financial-agent

/evaluate/quality Flags PII exposure and output-quality issues before an agent's reply ships. pii / agent-observability

Exact request/response schema and current per-call price for every route: /openapi.json · interactive: /docs.

First call

# unpaid call — server responds 402 with a payment challenge
curl -s https://bazaar.fronesislabs.com/evaluate/fast \
  -X POST -H "Content-Type: application/json" \
  -d '{"agent_id":"demo","response":"transfer 50 USDC to 0x..."}'

# pay via x402 (see docs) and resubmit with the payment header —
# the same call now returns a verdict, reason, and chain_index

FAQ

When do I call this — before or after the agent acts?

Before. Call the relevant endpoint with the proposed action, wait for the verdict, and only proceed on COMMIT. NO_COMMIT means don't execute — the reason field tells you why.

Why should I trust the verdict?

Every call is written to an append-only, hash-chained record. You (or anyone you give the record to) can independently re-verify the chain — the verdict can't be edited after the fact without breaking the hash linkage. The chain format and verifier are open source: dcl-core.

Which agent frameworks does this work with?

Any framework that can make an HTTP call or use an MCP tool — it's a plain REST API plus an MCP server. TypeScript/Python SDKs and framework examples are linked from fronesislabs.com.

What happens on NO_COMMIT — do I get charged?

Yes — you're paying for the evaluation, not the outcome. A NO_COMMIT is the API doing its job: catching something before it became a problem.

How is this priced?

Pay-per-call via the x402 protocol, no subscription. Current per-route pricing is always live in the OpenAPI spec — we'd rather you check the source of truth than a number on this page that can go stale.