Lyhna - AI Work Receipts your clients can trust

AI Work Receipts your clients can trust.

Before your AI tells a client "done," get a receipt that shows what it claimed, what actually crossed the tool boundary, what needs review, and what is safe to continue.

One product, several front doors. The proof layer stays underneath.

Client Review AI Work Receipt

Needs Review / DO NOT SEND

2 of 3 steps witnessed and supported. 1 claimed action the witness never saw - do not send.

Objective: Fix checkout rounding and confirm the fix with the client.
Supported: filesystem.write_file - call returned

test_runner.run_tests - call returned
Flagged: gmail.send - agent claimed it, witness saw no tool call
Next action: Confirm the client email happened before telling anyone it is done.

Click once. Watch the run. Get the receipt.

The public demo replays a committed receipt produced by the Lyhna witness loop with demo tools. It is not pretending to touch your inbox. It shows the product promise: the receipt separates witnessed work from agent self-report.

Lead with the pain your buyer already feels.

Agencies, operators, bookkeepers, and AI workflow builders do not need a lecture on infrastructure. They need a simple object they can hand to a client or teammate: here is what the AI claimed, here is what Lyhna witnessed, and here is what should not be sent yet.

Client AI Work Receipts

For agencies. Send a client-ready receipt after AI work instead of a fragile summary.

Do Not Send Checker

For deliverables. Catch "I emailed it" or "I shared it" when no send/share call was witnessed.

Workflow Path Audit

For automators. See whether the agent used the path it claimed, including wrappers and fallbacks.

Ops and Bookkeeping Handoff

For client services. Mark what is supported, what needs review, and what can continue tomorrow.

OKF Portable Export

For agent teams. Carry witnessed work into second brains, repos, and context systems.

The agent does the work. Lyhna witnesses the path.

Record the claim

The agent says what it believes happened: wrote the file, ran the test, sent the email, shared the doc.

Compare the actual call

Lyhna uses the witnessed tool-call path: system, action, result, and whether any evidence exists at all.

Print the receipt

The handoff labels each step as supported, mismatched, unsupported, do-not-send, or needs human approval.

The receipt is useful because it refuses to overclaim.

What Lyhna can say

This tool call was witnessed.
This claim matched the witnessed action.
This claim had no witnessed evidence.
This route differed from what the agent claimed.
This handoff is safe or not safe to continue.

What Lyhna cannot say

The client read the email.
The document is legally or business correct.
Every sentence the agent wrote is true.
Anything happened outside the observed workflow.
Your entire AI system is safe forever.

Proof stays underneath. The buyer sees the receipt.

Lyhna still has the technical spine: deterministic labels, proof refs, result hashes, offline verification, and portable OKF bundles. That is the floor under the product. The front door stays simple: what your AI claimed, what Lyhna witnessed, and what cannot be trusted yet.

Deterministic labels

No model opinion decides supported, unsupported, mismatch, or do-not-send. The labels come from rules.

Portable handoff

Receipts can travel as markdown, JSON, next-agent prompts, and OKF-compatible bundles.

Proof one click away

Developers can inspect receipts, hashes, and verification artifacts without making that the sales pitch.