Install · connect your agent

Connect your agent through the Lyhna MCP adapter.

Lyhna rides in the tool path your agent already uses, witnesses the calls, and renders the AI Work Receipt — locally, with your own keys, models, and tools. Here's how to wire it up, and exactly what works today.

What's real today. The Lyhna MCP adapter, the receipt renderer, and the capsule exports already exist in this repo — you can connect an agent and render receipts right now. The adapter installs from npm; the renderer runs from a source clone. Lyhna Desktop is the product direction: it packages this receipt layer into a local app with a receipt inbox. The desktop app isn't a download yet, and there's no one-command install or plugin-store listing — this page is honest about each step.

Install by agent

Paste this into your MCP-capable agent (Claude Code, Cursor, Codex…). It shows you a receipt now, connects your agent through the Lyhna MCP adapter so your calls are witnessed, and points to the guided flow for your first claimed-vs-witnessed capsule. It's a multi-step setup — your agent will stop and ask you for anything it can't do itself.

Help me set up Lyhna so I can see an AI Work Receipt for my agent's work. Lyhna is an MCP
adapter that witnesses the tool calls routed through it, locally, using my own keys and tools.
This is a guided, multi-step setup: do what you can in my terminal, and STOP to ask me for
anything you can't (a client restart, a signing key). Never fake a result or claim a step
worked if it didn't.

STEP 1 — Show me a real receipt right now (no setup, works offline):
   git clone https://github.com/Lyhna-ai/lyhna-witness && cd lyhna-witness
   node src/cli.mjs demo/live-loop-witness-input.json ./receipt --okf --pam
   Open ./receipt/CAPSULE.md and walk me through the verdict — that's the capsule shape.

STEP 2 — Connect MY agent through the adapter so its calls are witnessed (stdio):
   Add a server to my MCP client config (.mcp.json) that wraps an MCP server I already use.
   command = "npx", args = ["-y","@lyhna/mcp","stdio"], and point it at my server with
   LYHNA_PROXY_UPSTREAM_COMMAND / LYHNA_PROXY_UPSTREAM_ARGS_JSON, plus LYHNA_PROXY_LOOP_ID and
   LYHNA_PROXY_GOAL so the calls form a sealed loop (without those two the adapter runs with the loop
   disabled). LYHNA_PROXY_BIND_MODE=demo needs no key and makes no Lyhna-hosted call. Then tell me to
   reload my MCP client. Now my calls to that server route through Lyhna and earn a sealed receipt chain.
   (This connects and witnesses. It does NOT by itself produce the rendered
   claimed-vs-witnessed capsule — that's Step 3.)

STEP 3 — My first claimed-vs-witnessed capsule (guided, standing service):
   Comparing what I CLAIMED to what was WITNESSED needs the adapter's standing service: claim
   capture (LYHNA_PROXY_CLAIM_CAPTURE=1) plus a control loop you open and close. Follow the
   adapter's QUICKSTART "Path B": start the standing adapter, open a loop, record a claim
   (record_claim) and make a tool call, close/seal the loop, then export-pack to get a
   witness-input.json. Render it with:
       node src/cli.mjs <witness-input.json> ./receipt --okf --pam
   Use a signing key (LYHNA_API_KEY) for signed receipts; demo mode works unsigned with no key. Tell
   me exactly which of these you need me to do.

Two surfaces, on purpose: quick-connect (Step 2) witnesses your calls; the rendered claimed-vs-witnessed capsule (Step 3) is the adapter's standing-service flow — the guided QUICKSTART “Path B”. Lyhna only witnesses calls routed through it; anything else isn't on the receipt.

What actually works today

How you'll know it's working

Four states tell you exactly where you are.

Where this is going

The fastest way to judge Lyhna is to see the proof yourself: run the demo, copy the receipt, and audit it with your own AI. The same loop runs against your real tool calls today through the MCP adapter. Lyhna Desktop packages that into a local app — start the adapter, route your agents, and watch receipts land in a local inbox you own.

Lyhna Desktop is buy once, use with all your agents — see the ownership model. No numbers are published yet, and there's nothing to buy on this page.