Product
Send your agents' actions through AgentBlackBox, or connect the apps they work in. Rules decide what may happen, a person approves the risky steps, and every step joins a record you can verify and export.
AgentBlackBox sees and controls only what your agents send through it or through tools you connect. It does not monitor other AI tools.
Screenshots of AgentBlackBox with example data.



A record of the activity your agents report or route through AgentBlackBox, and of what happens in the apps you connect.
Agents send what they observed, decided and did, through the API or the SDK. Each task gets its own timeline.
Slack, Gmail, Google Drive, GitHub and Notion, connected with their own sign-in. Activity from them joins the record.
Each agent has a name, a risk class and the tools it may use. An action outside its tools is refused; an agent can be switched off.
Decisions, approvals and blocked actions over 7, 30 or 90 days, with a signed export.
Rules check every action your agents send through AgentBlackBox, before it runs.
Allow, ask a person first, or block, by action type, risk, amount, recipient, tool or content. Rules can also change or redact an action. Built-in rules already block deletions and ask before money moves or data leaves your organization.
New workspaces start in watch-only mode: rules decide and record, nothing is blocked. Switch to enforce when the decisions look right.
AgentBlackBox sets the risk itself: sending outside your organization or moving money raises it, and an agent can never lower it.
Actions go through three gateways: agent decisions, connected-app actions and custom tools. A held action runs only after it is approved. App credentials stay encrypted with AgentBlackBox; the agent never holds them.
Every rule change is a new version you can roll back. "Test a rule" shows the decision for a sample action without recording it.
Each agent key has limits on the gateways. If a limit cannot be checked, the action is refused rather than run unchecked.
Risky actions wait for a human decision, recorded with the action and its risk.
In a team, the approver must be someone other than the person or agent that asked. Approver groups decide who may approve what.
Working alone? You confirm your own risky requests by entering your password again, and the record notes it.
A request expires after 72 hours without a decision, and approving checks the action against the rules in force at that moment.
Multi-step workflows with an approval step, run on demand.
Workflows on a schedule, reminders for waiting approvals and escalation to a group's backup approver.
A tamper-evident record, and exports for auditors and clients.
Every recorded event is hashed with SHA-256 and chained to the one before it. Verifying shows whether a recorded event was changed or removed from the middle of the chain.
Signed checkpoints seal a range of events, so removing the most recent ones is detected too. Admins create them; automatic checkpoints need scheduled jobs.
A JSON export signed by AgentBlackBox (HMAC-SHA256) and a printable audit report for every task; owners can export the whole workspace. Checking a signature needs AgentBlackBox's key.
Who changed which setting, rule, key or member, and when.
Anchoring checkpoint hashes to a public timestamp service (OpenTimestamps).
Public-key signatures, so a third party can check an export without our key.
Add a policy check and an approval step to your agent with a few API calls.
Record events, evaluate rules, request and poll approvals, and send actions through the gateways. Every answer carries a request id.
Personal or service keys, optionally bound to one agent and given an expiry. Shown once and stored only as hashes.
Register your own HTTP endpoints and MCP servers; agents call them through the same rules. Only public https addresses are accepted.
Inbound webhooks are signed with a timestamp and refused when replayed.
A client for every endpoint, including the modify, redact and retry answers. Available in the repository; not published as a package yet.
Roles, company sign-in and the controls a security team asks for.
Owner, admin and member. Owners manage owners and role settings; admins can deactivate members; a workspace always keeps an active owner.
Microsoft Entra sign-in with proof that your organization owns its email domain, and automatic member provisioning.
Keep all records under legal hold; export or erase a person's data.
Stream the audit log to Splunk, Datadog or a webhook, and notify Slack channels.
Sign-in and recovery.
Passwords stored as bcrypt hashes; password sign-in pauses after 10 wrong attempts; changing your password or "Sign out everywhere" ends every session.
Single-use links that expire. Until email sending is switched on, support helps you back in.
A second factor for password accounts.
Personal, Pro and Business subscriptions through Stripe (coming soon, billed monthly). Every feature works on the free plan today.