Start with an owner who prepares the audit and an independent reviewer who approves and concludes it. Link this login to the person using it.
Each person needs their own linked login. Administrators can create or link the second person's login under People & Authority.
An agent represents a service or bot that sends requests through the gateway.
The gateway injects your key into requests so agents never see it directly.
Point your agent's API calls at the gateway. Every request and response gets recorded.
Are you sure?
AI audit workspace. Sign in to continue.
Ask your administrator for an access credential. Administrators issue credentials under Administration → People & access. Keep your credential private.
Audit posture at a glance: evidence coverage, record integrity, control effectiveness and environmental impact across all AI systems in scope. Every figure is a view over the evidence record; nothing here can be edited.
| When | Agent | What happened | Outcome | Took | |
|---|---|---|---|---|---|
Calls appear here as soon as an agent sends one through the gateway. Help › Start here shows how to send a first call.
Review whether your AI controls are supported by evidence.
Define the systems, period and criteria. Assigned analysts review the evidence before an independent reviewer approves the conclusion.
| Name | Period | Status | Owner | Created | Actions |
|---|---|---|---|---|---|
Create an engagement to define which AI systems, agents, providers and time period you are auditing. Engagements, control tests and findings will be managed here.
Every AI interaction recorded in the evidence ledger, newest first. Open a row to read the conversation, verify the record's integrity, and trace the decision lineage across protocols.
| Time | Agent & session | Action | Result | Controlled by | Duration | Tokens | Record ids | |
|---|---|---|---|---|---|---|---|---|
Widen the time range or clear a filter. Calls appear here as soon as an agent sends one through the gateway.
Independent verification of evidence integrity. Each agent's records form a tamper-evident chain where every entry carries a fingerprint of the one before it. The gateway periodically signs a checkpoint over a batch of them. Opening this page does not run a full check. Verification checks stored record fingerprints, payloads, chain links and receipts; external witnessing is a separate protection.
Review stored records, gateway signing evidence and external publication separately.
The gateway signs a checkpoint once enough records have accumulated, or within an hour of the oldest unsigned record. Nothing to do; one will appear.
| Agent | Status | Records | Integrity | Latest fingerprint | Started | Actions |
|---|---|---|---|---|---|---|
A per-agent record starts with the agent's first call. Each event carries a fingerprint of the one before it, so any change or gap is detectable.
Human oversight and accountability. When a control policy requires human authorization, requests are held here until qualified people sign them. People, their roles and access tokens are managed below.
| Decision | AI system | Due | Review |
|---|---|---|---|
When an AI system has an approval policy, its requests are held here until named people approve them. Configure the rule under Policies.
| Name | Identifier | Role | Status | Record id | Actions |
|---|---|---|---|---|---|
Add the reviewers and approvers who may sign held requests, then link their access tokens to them below.
| Label | Can do | Signs as | Status | Created | Token id | Link to a person |
|---|
Create a token to sign in to this dashboard or to call the API. Link it to a person if it should be able to approve requests.
Track issues until someone resolves and reviews them.
Open a finding’s audit to review its evidence, owner and next action.
Findings are created from control tests within an audit engagement. Create an engagement under Audits, then add control tests and document findings as you review the evidence.
| Issue | Audit | Why it matters / next step | Severity | Status | Action |
|---|
Sustainability audit evidence: estimated environmental impact of every AI interaction, recorded as append-only receipts with published methodology. Every figure is an estimated range; what is verifiable is that each estimate was recorded with its call and never changed.
No environmental budgets set. Budgets are set per agent through the API.
| Agent | Events | Energy (mWh mid) | Water (uL mid) | Carbon (mgCO2e mid) | Evidence |
|---|---|---|---|---|---|
Every completed model call gets an estimate automatically. All figures are estimated ranges, never measurements.
| When | Agent | Model | Tokens (weighted) | Energy (likely) | Evidence quality | Origin | Receipt id · position · fingerprint |
|---|
No receipts yet. One is written with every completed model call.
| Agent | Level | Efficient Models | Routine Keywords |
|---|
No recommendations configured. They are set per agent through the API and only recorded, never applied.
| Agent | Energy Limit | Water Limit | Carbon Limit | Period | Action |
|---|
No environmental budgets set.
AI systems, collection planes and evidence infrastructure. Register agents, provider keys, MCP servers, A2A destinations, signing keys and retention policies that define how AI activity is captured.
| Name | Where it runs | Records | Status | Kept for | Registered | Agent id | Actions |
|---|---|---|---|---|---|---|---|
An agent is a bot or service that sends its AI calls through the gateway. Register one and you will get a token to give it.
Choose whose model connection to inspect. Subscription and organization API credentials use separate registrations.
Origin only. Select one agent; private endpoint keys cannot apply to all agents.
| Provider | Label | Used by | Status | Health | Added | Key id | Actions |
|---|---|---|---|---|---|---|---|
Add an Anthropic, OpenAI or Google API key. The gateway adds it to each call on the agent's behalf; agents never see it.
| Name | Endpoint | Used by | Auth | Trust | Status | Added | Server id | Actions |
|---|---|---|---|---|---|---|---|---|
Register an MCP server so agents can call its tools and resources through the gateway. The gateway holds the server credential; agents never see it. An agent then calls it by name:
curl -X POST http://GATEWAY/proxy/mcp/SERVER_NAME \
-H "X-Gateway-Token: AGENT_TOKEN" -H "Content-Type: application/json" \
-d '{"jsonrpc":"2.0","id":1,"method":"tools/call","params":{"name":"TOOL","arguments":{}}}'
| Name | Endpoint | Used by | Auth | Max depth | Protocol | Card | Status | Added | Dest id | Actions |
|---|---|---|---|---|---|---|---|---|---|---|
Register a remote agent so your agents can delegate tasks to it. The gateway records every delegation, what came back, and refuses a delegation deeper than the limit you set. An agent then delegates by name:
curl -X POST http://GATEWAY/proxy/a2a/DESTINATION_NAME \
-H "X-Gateway-Token: AGENT_TOKEN" -H "Content-Type: application/json" \
-d '{"jsonrpc":"2.0","id":1,"method":"message/send","params":{"message":{"messageId":"m1","role":"user","parts":[{"kind":"text","text":"Summarise Q3"}]}}}'
| Name | Local agent | Relays into | Card | Default | Status | Added | Front id | Actions |
|---|---|---|---|---|---|---|---|---|
A front lets a remote peer send tasks to one of your agents through this gateway. Register the agent's own A2A endpoint as a destination first, then add a front over it. The gateway publishes a signed agent card for the front, checks the peer's token and delegation token, records every task received, and relays it to the agent.
| Name | May reach | Keys | Delegation token | Last seen | Status | Added | Peer id | Actions |
|---|---|---|---|---|---|---|---|---|
A peer is a remote gateway or agent allowed to send tasks to your fronts. Adding one issues a token, shown once, that the peer presents as a bearer token. Give a peer gateway's key set URL so its signed delegation tokens can be verified and the depth of a chain carries across.
curl -X POST http://GATEWAY/a2a/FRONT_NAME \
-H "Authorization: Bearer PEER_TOKEN" -H "Content-Type: application/json" \
-d '{"jsonrpc":"2.0","id":1,"method":"SendMessage","params":{"message":{"messageId":"m1","role":"ROLE_USER","parts":[{"text":"Summarise Q3"}]}}}'
| AI system | Actions held | Authorization | Deadline | Updated | Actions |
|---|---|---|---|---|---|
Add a policy when an AI action must stop at the gateway until an accountable person reviews the captured request.
| Applies to | Kept for | Then | Updated |
|---|
Retention says how long the content of records is kept before it is redacted. Rules are set per agent through the API; the license sets the default.
The gateway can send each event to an endpoint of yours (a SIEM, a ticketing system) with a signed payload and retries. Subscriptions are created and inspected through the API; see the Webhook Reference in Help.
How to use the audit workspace end to end, what it records, and what the terms on these screens mean.
curl -X POST http://GATEWAY/proxy/anthropic/v1/messages \
-H "X-Gateway-Token: AGENT_TOKEN" -H "Content-Type: application/json" \
-d '{"model":"claude-sonnet-4-20250514","max_tokens":50,"messages":[{"role":"user","content":"Say hello"}]}'
The provider's answer comes back to you unchanged./proxy/mcp/NAME or /proxy/a2a/NAME with the same agent token. The gateway records the request, the reply, and whether the tool or task actually succeeded: a tool that reports isError and a task the remote agent reports as failed both show as Failed. Add X-Gateway-Parent-Event-Id to a later model call and "Show lineage" links them.--attest-every 1) Integrity shows a signed checkpoint that includes the call and its receipt; the receipt's "Inclusion proof" button verifies it against the signature.Recorded: every model call an agent sends through the gateway (the request and the answer, as sent and received); every tool or API call an agent routes through the gateway's tool proxy; tool use that happens inside a model answer (the model asking for a tool, the agent returning the result), because it is part of the recorded conversation; and anything an agent chooses to report about its own local activity (file reads, local tool runs, decisions).
Not recorded: anything an agent does without going through the gateway. Self-reported activity is taken at the agent's word; the gateway cannot check it. The evidence shows what passed through the gateway, not everything the agent did.
What "verified" means: every record carries a fingerprint of its content and of the record before it. Verification recomputes those fingerprints and compares them with what is stored, and checks the signatures on the checkpoints. A pass means nothing was changed, inserted or removed since it was recorded. It does not mean the content is true, and environmental figures are always estimates.
The full documentation set ships with the gateway in the microbiocol/ folder. The ones a reviewer needs first: the QA Walkthrough (38), the Auditor Guide (10), the Evidence Package Verification Specification (18), the Environmental Methodology Specification (34) and the API Reference (03).
License, plan limits and workspace administration.
Admin, employee and auditor accounts are included on every tier. Upgrade to Professional for unlimited agents, webhooks, and 7-year evidence retention.
View Plans