For LangGraph agents · docketproof.dev

Put your agent on the record.

DocketProof pauses risky tool calls for approval before they run — then files a signed, chained entry for every decision. Anyone can verify the docket later, offline, without trusting us.

Every entry is evidence — approved or denied.

$ pip install docketproof
See how it works →

Exhibit A

Tamper with one entry. Watch the whole docket after it fall.

This is a real, in-browser SHA-256 hash chain — not an animation. Each entry references the hash of the one before it, the same principle the real docket uses with ed25519 signatures. Alter entry № 0002 below and re-verify: notice that entry itself fails, and every entry filed after it becomes unverifiable too — even though nothing about them changed. That's how a chain of independently signed events works.

THE DOCKET · 5 entries, chain intact

Click Edit entry № 0002 — notice it lights up blue immediately, since something changed. Then click Re-verify the docket and watch that entry, and everything after it, turn red. You can also click directly into any entry's resource_id and type something else. Nothing leaves this tab — every hash is computed locally with the browser's own SHA-256. The full docket uses ed25519 signatures over an RFC 8785-canonicalized payload; see the format spec for the exact bytes.

Procedure

Four steps, all on your machine.

No hosted service is required for the core mechanism. Nothing about your agent's data leaves your process unless you turn on the hosted dashboard.

i.

Agent tries something risky

Deleting a resource, sending a message, making a payment — anything you've marked as needing approval.

ii.

Execution stops right there

The agent waits at exactly that point. Nothing runs until a decision is made.

iii.

The decision is filed

Allow, deny, or hold for review. Both outcomes get signed — a denial is evidence too.

iv.

Anyone verifies, offline

ed25519 signature, hash-chained. Verify with the open CLI — no account, no API key.

Filing

One line added. No graph rewrite.

agent.py 
# pip install docketproof

from docketproof import guarded_tool, Policy

policy = Policy.from_file("policy.yaml")

@guarded_tool(policy=policy, docket_path="./docket.jsonl")
def delete_resource(resource_id: str) -> str:
    # your existing tool, unchanged
    ...

Scope of the record

What an entry proves — and what it doesn't.

✓ What it proves

  • This exact tool call, with this exact input, was evaluated at this exact time.
  • The named policy produced this named verdict — allow, deny, or review.
  • If allowed, this is the exact call that proceeded — nothing was altered after filing.
  • The entry has not been modified since it was signed. Any edit breaks verification.

✗ What it doesn't prove

  • That the policy itself was correct or well-designed.
  • That the agent's underlying intent was safe, legal, or sensible.
  • Anything about actions outside this specific docket.
  • That the person who approved it should have.

Why now

This isn't a hypothetical problem.

Scale and regulation are moving in the same direction at the same time.

3M+

AI agents are operating inside companies today — under half of them are monitored in any way.

Gravitee, State of AI Agent Security 2026

88%

of organizations reported a confirmed or suspected AI agent security incident in the past year.

Gravitee, State of AI Agent Security 2026

Dec 2, 2027

deadline for the EU's full high-risk regime. The AI literacy duty (Art. 4) and transparency rules (Art. 50) already apply today.

Regulation (EU) 2026/1744

Get in touch → [email protected]

Terms

Free without limits if you self-host. Pay only for the hosted convenience.

The SDK, the policy engine, and the verifier are open source and MIT-licensed — that stays true regardless of plan, and self-hosted really is unlimited: nothing touches our servers, so there's no artificial cap on the free tier like you'll find on hosted-only tools. You pay for entries synced to the hosted dashboard — not seats, not local usage. The hosted dashboard runs on EU-based infrastructure — docket data for hosted plans does not leave the EU, which matters if you're building evidence for EU data-residency or AI Act requirements.

Open Source
€0

Self-hosted, genuinely unlimited — it never reaches our infrastructure.

  • pip install docketproof
  • Local policy engine, no usage cap
  • Signed, chained entries on disk
  • Independent open-source verifier
  • Community support (GitHub issues)
pip install →
Team most teams start here
€29/month

5,000 hosted entries/mo included, then €2.50 per 1,000.

  • Everything in Open Source
  • Hosted dashboard for the docket
  • Slack approval routing
  • Shared team policies
  • 90-day retention + export
Start 14-day trial →
Business
€199/month

50,000 hosted entries/mo included, then €1.50 per 1,000.

  • Everything in Team
  • SSO
  • Unlimited retention
  • Signed audit export bundles
  • Priority support
Start 14-day trial →

Enterprise (on-prem/VPC, dedicated support): custom pricing. Prices in EUR.