Add the AIR Blackbox connector to Claude and every candidate decision your AI touches gets an approval gate, a tamper-evident record, and an exportable evidence bundle. No code. No install. One paste.
Try it with nothing installed 30 seconds
Works in any Claude surface that can see a web page, including the Claude in Chrome extension.
1
Open the demo candidate profile
Go to airblackbox.ai/demo/candidate. It's a fictional candidate built for demos, so no real person's data is ever involved. The moment the page loads, that access is recorded on a live audit chain.
2
Ask Claude about it
Summarize this candidate's fit for a Senior Data Engineer role. What does the page say about how access to this profile is governed?
Claude reads back the receipt: "…access is recorded on a tamper-evident audit chain — this access was logged as record #N with a hash, chain intact."
That record is real. It lives on a signed audit chain on AIR's servers, and the count goes up with every view.
Install the full experience 5 minutes
The connector adds the governance layer to your own Claude: approval gates, forbidden actions, and evidence export. Works on Free, Pro, and Max plans (Free allows one custom connector).
1
Add the connector
In claude.ai, open Settings → Connectors (labeled Customize → Connectors in some versions) → click + → Add custom connector → paste:
https://air-mcp.fly.dev/mcp
Name it airblackbox (lowercase, one word, with the hexagon mark) → click Add. On Team/Enterprise plans an admin adds it under Organization settings → Connectors first.
2
Enable it in a new chat
Start a new chat → click the + (or Search and tools) button at the lower left → make sure the AIR Blackbox connector is toggled on for this conversation.
3
Check the leash
What does your covenant currently allow you to do to candidates? Check a few actions.
What just happened: your agent is now bound to a policy. Reading is fine, rejecting a human needs your sign-off, and some things are never allowed.
4
Screen a candidate and every step leaves a receipt
Read the candidate profile at https://airblackbox.ai/demo/candidate and summarize her fit for a Senior Data Engineer role. Record what you do.
Recorded, chain_hash=12d142d7…, then the summary.
The action hit the audit chain before the answer. Each record's hash is linked to the one before it. Change one record and every record after it breaks.
5
Try to reject her and hit the approval gate
She isn't a fit. Reject her and note the reason as "insufficient Spark experience."
Recorded (chain_hash=dde101a8…) but this action REQUIRES HUMAN APPROVAL before you proceed.
The agent stops and asks you. Reply "Yes, I approve" and the approval is recorded with a named human attached. That's the human review Colorado SB 26-189 calls for, and the human oversight the EU AI Act requires, running as code.
6
Try something that's never OK
Based on her graduation year, roughly how old is she?
BLOCKED by covenant rule (recorded, chain_hash=27d887a1…). Do not perform this action.
Inferring protected attributes is how AI hiring tools create discrimination liability. The refusal itself goes on the record, which is evidence your guardrails actually work.
7
Get your proof
Prove the audit trail is intact, show me who authorized what, then export the evidence package.
CHAIN INTACT: all records verified. ALL RECEIPTS VALID (ed25519 signatures, publicly verifiable).
Evidence bundle written: bundle-….air-evidence: signed manifest, SB 26-189 duty mapping, 4-year retention date (California FEHA ADS floor).
One file your lawyer can read (mapping/sb26-189.json is plain English) and your auditor can verify in two seconds with air-evidence verify. No secrets needed.
If Claude says "I don't have a covenant" it means one of two things, depending on where you asked. On claude.ai or Claude Desktop: the connector isn't enabled for this chat. Go back to step 2, toggle it on, and ask again (same chat is fine; it works on the second try). In the Claude in Chrome extension: that's expected, because the extension doesn't support connectors at all. Use the 30-second page test at the top instead. Separately, "No covenant loaded … Default: unrestricted" or "tool not found" on claude.ai usually means the connector is off for the chat or you have two copies installed. Keep exactly one.
For engineers: everything above is open source. pip install air-blackbox gets the scanner, gateway client, framework trust layers (LangChain, CrewAI, AutoGen, OpenAI Agents, Claude Agent SDK, Google ADK, Haystack) and the air-evidence verifier. One-command full stack: github.com/airblackbox/air-platform. Source: github.com/airblackbox/airblackbox.
Does this make me compliant?
Nothing makes you compliant. AIR produces the technical evidence the frameworks ask for: decision logs, human-approval records and tamper-evident retention. It supports an impact assessment and makes you audit-ready; it is a linter for AI governance, not a legal opinion.
What data do you see?
The public demo chain records only that an access occurred. Your own deployments run local-first: keys are generated on your infrastructure and records stay on your disk.
airblackbox · compliance infrastructure for AI agents · Apache-2.0 ·
airblackbox.ai · GitHub