Wrap your agent in a single API call. AgentWall checks every prompt for PII and prompt injection, restricts every tool call to the destinations you approve, and writes a tamper-evident log your auditor can use as evidence. Model-agnostic. Provider-agnostic. Pay only for what you inspect.
Deterministic detectors on every prompt, every tool call, and every response. No fuzzy ML scores — clear rules with explainable verdicts.
Catch personal data on its way into the model and on its way out. Block, redact, or just log — the choice is yours per category.
Stop API keys and credentials from ever reaching the model — or worse, ending up in a chat response your user copy-pastes into a Slack thread.
Block role overrides, jailbreaks, and exfiltration attempts before they reach your model. Deterministic patterns, sub-millisecond latency.
Your agent can call `send_email` — but only to your domains. It can call `fetch` — but only to your API. Hard caps stop runaway loops.
Verdicts are only useful if you can debug them. Replay any run, see why every rule fired, and verify your guardrails are still catching attacks today.
Pass a `run_id` on every inspect call, and you can replay any agent run frame-by-frame from the dashboard. Every input, tool call, output, and verdict on one timeline.
GET /api/dashboard/logs/replay/<run_id>Every block, redact, or allow shows the exact rule that fired and the substring it matched. No black boxes, no confidence-score guessing.
Every /v1/inspect response — findings[].pattern_id + matchedEvery decision your agent makes — and every verdict we give it — recorded to disk with fsync. Exportable on demand. The kind of log auditors actually accept.
A live "guardrails verified" check that runs eight known-bad payloads through your policy on demand. Proves your protections are still operating today, not the day you set them up.
POST /api/dashboard/canary/runAdopt without breaking production. Bring your own model. Push events anywhere. AgentWall sits in the middle of your agent stack and stays out of the way.
Drop AgentWall in front of any agent and watch — without blocking a single request. Build a behavioral profile, generate a policy from real traffic, switch to live when you trust it.
Stop your model from inventing facts it couldn’t have known. A fast heuristic check runs on every output; smart mode uses an LLM for deeper verification.
The LLM-powered checks (hallucination, semantic compression) can run on your own provider. Your key, your bill, your data path — anything OpenAI-compatible.
Push verdicts and account events to your own stack the moment they happen. Slack, PagerDuty, your own backend — anywhere that speaks HTTP.
Made for teams shipping in regulated industries. The evidence your auditor asks for, the privacy endpoints your users have a right to, the deletion controls your DPO requires — all included.
One-click download of a ZIP containing seven CSVs, each mapped to a SOC 2 Trust Services Criteria control, plus a README for your auditor. Generated live from your audit log.
GET /api/dashboard/compliance/soc2-packA privacy-preserving endpoint your users can hit through your own dashboard. Shows them what happened to their data — counts, categories, timestamps, never content.
GET /v1/transparency?user_id=<id>When a user requests deletion, your data goes with them. AgentWall anonymises the user record and purges API keys, transactions, and usage in one transaction.