The accountability layer for AI agents
We gave agents power. We forgot accountability.
Agents now run for hours inside your databases, APIs, and financial systems — and the only record of what they did is a log they can edit themselves. Merkl turns every agent action into a tamper-proof, independently verifiable record. When someone asks what your agent did, you don't point to your own logs. You prove it.
Every tool call, prompt, and approval — recorded the moment it happens.
The blind spot
Six hours. Zero accountability.
An agent triages a patient. It reads the chart, scores the symptoms, assigns priority 3. Hours later: "Task completed." Someone dies. In the courtroom, who can prove what the agent actually did? Today, the honest answer is nobody.
5h 47m of unverifiable activity
Observability tells you what happened. Accountability proves it.
01 · The record
An audit that can't be gamed.
An audit works because there's a record on file the company can't go back and edit. Merkl is that record — for your agent. Every action is fingerprinted the moment it happens and folded into a single proof. You keep your own copy of what happened; the record on file is the independent truth. When anyone checks, the two reconcile — or the tampering shows.
This is an inclusion proof.
02 · Append-only
The audited can't edit the audit.
Each session joins an append-only record the agent and its platform can't touch. Every entry is bound to the one before it — remove or reorder anything, and it no longer adds up, in plain sight.
Reorder one entry and the chain breaks.
03 · Permanence
Signed, and sealed beyond reach.
Merkl signs the record's head with a published key, then seals it into an independent public record. Once it lands, the moment is fixed — impossible to backdate or rewrite, by us or anyone.
Fixed. Immovable.
Open verification
Verify it yourself. Trust no one.
Every proof is a standalone page that checks itself in your browser — offline, no account, nothing installed. If your audit trail needs you to trust the vendor, it hasn't solved the trust problem. Anyone can re-check any action against the record; change one byte and it says so.
Read the codeTwo lines. Any agent.
Merkl is a decorator, not a rewrite. Wrap your agent and every action is recorded — the SDK does the fingerprinting on your machine, and the notary only ever sees hashes and tool names, never your data. Works with Claude Code today; LangChain, OpenAI, CrewAI and more via the SDK.
Accountability today. Enforceable guardrails next.
$ pip install merkl-sdk
$ merkl install --claude-code --global