AI agent accountability

Every AI agent action, logged and hash-chained.

Tamper one row and the chain breaks visibly. An autonomous tribunal audits the log and exports a regulator-ready evidence pack.

Get started free View on GitHub

demo-session.chain

chain: INTACT

Click an event to edit it. The same verifyChain code that /verify runs checks the chain in your browser.

The EU AI Act asks for proof, not just logs.

€35M

Maximum fine for the most serious violations (prohibited AI practices, Art.5), or 7% of global revenue. Other obligations carry lower tiers.

2026

Article 50 transparency obligations (chatbot and synthetic-media disclosure) take effect. Full high-risk obligations are delayed to December 2027 under the Digital Omnibus.

Art.15

Robustness and cybersecurity requirements your agent logs must now demonstrate.

Four steps from first action to court-ready evidence.

Connect

Install the recorder from the GitHub repo (pip install -e .), then add it to Claude Desktop or Claude Code as an MCP server. No code in the agent. Any other agent can POST to /events with an API key.

{
  "mcpServers": {
    "teluvane": {
      "command": "teluvane-mcp",
      "env": {
        "TELUVANE_URL": "https://blackbox-agent-accountability.onrender.com",
        "TELUVANE_API_KEY": "tv_live_..."
      }
    }
  }
}

Recorder

Every agent action, LLM call, tool invocation, and result is appended to a SHA-256 hash-chained log. Any silent edit breaks the chain when it is next verified.

demo session · event #2 · tool_call
action: send_email
hash: bccf52...1f01
prev: 5b60a3...db7c

Tribunal

An LLM auditor checks the full log against each rule in a policy pack (EU AI Act, ISO 42001, NIST AI RMF, or SOC 2) and cites the events and framework reference behind every finding. Without an API key, a deterministic offline detector runs instead.

rule: data_exfiltration
severity: critical
ref: Art.12 record-keeping; Art.15 robustness

Evidence Pack

One click exports a report with the violation table, the full action log, and the chain-integrity status. HTML on every plan, PDF on paid plans.

evidence pack · demo-session
3 events
chain: INTACT

Verification walks the whole chain, not just the last row.

Each event stores the hash of the one before it. Change a single byte in any event and verification fails at that event, and nothing after it can be trusted. Run it on a session from the dashboard, or edit an event in the card above and watch it happen.

verify(chain) → 3/3 events valid (demo session) · INTACT

We're early-stage. Here's what's already true.

No SOC 2 report yet, no formal certification. Rather than a badge we haven't earned, here's how the system is actually built.

Org isolation is enforced in code, not just convention.

Every database query is scoped to an organization at the data-access layer. A query missing that scope throws before it runs, rather than depending on every developer remembering to filter correctly.

query.where(org_id=current_org.id) # required, or the query throws

Two separate credential paths.

Dashboard logins (Supabase; tokens are checked against the project's published signing keys, or its shared secret on older projects) and agent event ingestion (per-org API keys) never share credentials. A leaked dashboard session can't be used to forge log entries, and vice versa.

verify_jwt(session) != verify_api_key(org_key) # disjoint paths

The hash chain detects tampering, it doesn't prevent it.

A privileged database user can still edit a stored row. What the chain guarantees is that the edit becomes visible the next time anyone verifies the log, instead of staying silent.

sha256(event[i-1].hash + event[i].payload) == event[i].hash

Pricing

Free

Run it on one project. Offline detector, no live AI calls, no card required.

✓Unlimited manual audits (offline detector)
✓SHA-256 hash-chained event log
✓HTML evidence pack export
✓EU AI Act policy pack

Get started free

Starter

$9.99_/mo

Real AI-audited verdicts, not just keyword matching, for a small team getting started.

✓Everything in Free
✓Hosted Claude tribunal, up to 15 audited sessions/mo
✓PDF evidence pack export
✓Email support

Get started

Recommended

Pro

$19.99_/mo

Managed cloud. Everything you need for a production AI team, without running your own infrastructure.

✓Up to 10 agents managed
✓Hosted Claude tribunal, up to 50 audited sessions/mo
✓Automated tribunal runs on schedule
✓EU AI Act, ISO 42001, NIST AI RMF, and SOC 2 policy packs
✓PDF + HTML evidence pack exports
✓Custom policy rules
✓On-chain session anchoring (Avalanche)
✓Priority email support