# Agent Guard / Agent Trust Firewall — developer quickstart

Free advisory HTTP API and remote MCP. No payment, card or wallet. Public checks need no key. For measurable pilot usage, activate an email-verified identity at https://agent-trust-firewall.andrea-vignoni.workers.dev/trial?source=organic-docs and use authenticated routes.

## Choose your integration

- HTTP/OpenAPI: https://agent-trust-firewall.andrea-vignoni.workers.dev/openapi.json
- Remote MCP: https://agent-trust-firewall.andrea-vignoni.workers.dev/mcp
- Python example: /examples/guard.py (Python 3.10+, standard library only)
- TypeScript example: /examples/guard.ts (Node.js 22.18+; node guard.ts)
- MCP client config: /examples/mcp.json (syntax varies by client)

Run examples with no credentials to exercise public checks; this is a test, not a verified activation. Optionally set AGENT_GUARD_KEY to your activated beta key to exercise authenticated /trial/api/trust_check_spend. The key is read from the process environment and never printed. The example sends synthetic inputs only. Do not send secrets or customer data.

Both examples check amount=2 and amount=11 against budget=10. They expect ALLOW and BLOCK. A real application must execute its tool only after ok=true and verdict=ALLOW; errors, timeout, unknown verdict and BLOCK must stop execution. No example executes a payment. The check does not reserve budget: your application must manage concurrent spending and enforce its own limits.

## MCP lifecycle

POST JSON-RPC initialize (protocolVersion 2025-06-18), send notifications/initialized, then tools/list and tools/call. Use Accept: application/json, text/event-stream and Content-Type: application/json. Canonical tools: trust_check_action, trust_check_spend, trust_validate_data, trust_clean_data, trust_get_capabilities. A transport-level success or isError=false does not mean ALLOW: inspect structuredContent.verdict (or parse the text content). REVIEW requires inspection. The public MCP currently has no authenticated pilot metering; discovery probes and tool calls are not evidence of independent adoption.

## Attribution and future interest

The trial page discloses storage of an integration's identity ID, source label, dates, counters and optional interest. Choose a published source (glama, organic-docs, claudecodeagents). Changing a client header cannot establish independence. Authenticated usage is recorded server-side after input validation and quota reservation. Trial registrations, first call, usage on multiple days and voluntary pricing interest are separate measures. Owner tests are qa=1. Other registrations remain unreviewed until independent integration evidence is checked; email verification alone is insufficient.

Future interest is optional and withdrawable on /trial. Proposed EUR 9/month is research only: no subscription, debt, payment mandate or automatic billing. Do not count interest as revenue.

## Limits

64 KiB JSON request; data cleaner up to 500 rows. Stateless inline policies; unauthenticated source labels; deterministic structure checks, not factual verification. No wallet control or reliable prompt-injection prevention. Fail closed and handle 400/401/403/413/429/5xx. No automatic retry of an external side effect.

## Help / first external integration

Email andrea.vignoni82@gmail.com with subject 'Agent Guard independent integration'. Include your public project URL (if available), environment, one successful response and one enforced BLOCK, and whether the integration is independent of the maintainer. Redact keys and private data. Evidence is reviewed before counting adoption. The private development repository is not required for these public docs.