Run local AI.
Keep the receipts.

Install Free

INVAR installs beside the OS you already run — nothing is flashed or replaced. It creates a private environment under ~/.invar and puts one invar command on your PATH. Uninstall is rm -rf ~/.invar ~/.local/bin/invar.

curl -fsSL https://www.anomly.com/get/invar.sh | sh

Requirements: Python 3.10+ and a llama.cpp llama-cli on your PATH (or set INVAR_LLAMA_BIN). No llama.cpp yet? brew install llama.cpp on macOS, or grab a prebuilt binary from the llama.cpp releases.

Get a model

Any llama.cpp-compatible GGUF works. A small one to start: download the q8_0 file from SmolLM2-135M-Instruct-GGUF and save it under ~/.invar/models/. Or point INVAR at any model you already run — it wraps whatever llama.cpp can load.

Your first receipt

invar serve --model ~/.invar/models/<your-model>.gguf

# in another terminal:
curl -s localhost:8577/v1/chat/completions \
  -H 'Content-Type: application/json' \
  -d '{"messages":[{"role":"user","content":"The capital of France is"}]}'

The response is OpenAI-shaped plus a receipt: a SHA-256 certificate over the runtime binary, model weights, prompt, decode parameters, and output — hash-chained into worldline.jsonl.

Verify a receipt

invar verify worldline.jsonl --binary llama-cli --model ~/.invar/models/<your-model>.gguf

ACCEPT — re-executed, output digest matches means the pinned computation was re-run and produced the same output, bit for bit. Edit any byte of the log and verification REJECTS. The verifier is free for everyone, forever — a receipt nobody can check isn't a receipt.

The endpoint

invar serve is an OpenAI-compatible endpoint, so it drops into whatever client or agent framework you already use. It binds to 127.0.0.1 by default; expose it deliberately with --host behind your own TLS. Every completion carries its receipt in the response body.

Worldlines

A worldline is the append-only, hash-chained log of a deployment's receipts, rooted at a genesis hash. Each entry links to the previous one, so the whole history is tamper-evident: change or drop any entry and every later link breaks. Export or ship it wherever your records live.

Teams: the Ledger Licensed

The INVAR Ledger collects worldline entries from every device in a fleet, verifies each one at ingest, and exports certified chain-of-custody packets. It's the paid, self-hosted component — your data never touches Anomly.

Run the collector

INVAR_LICENSE=/etc/invar/license.invar \
LEDGER_DIR=/var/lib/invar-ledger \
LEDGER_TOKEN=$(openssl rand -hex 24) \
invar ledger

It refuses to start without a valid, unexpired license. Put a TLS reverse proxy (Caddy/nginx) in front if you serve a fleet, and set HOST=0.0.0.0.

Connect agents

LEDGER_URL=https://ledger.yourco.com LEDGER_TOKEN=... \
INVAR_DEVICE_ID=$(hostname) invar serve --model ...

Pushes are best-effort: a Ledger outage never blocks local inference, and the device's own worldline remains the source of truth. A device whose history was rewritten locally is refused at ingest — that refusal is itself the audit signal.

Export a chain of custody

curl -H "Authorization: Bearer $TOKEN" \
  "https://ledger.yourco.com/v1/export?device=alice-laptop" > custody.json

The packet contains the entries, a per-entry verification report, the collector attestation, and a certificate any third party can recompute with the free verifier — no Anomly involvement required.

Licensing

An INVAR license is an Ed25519-signed file, verified offline by the software against a public key it ships with. No activation server, no phone-home, no account.

invar license verify license.invar
# VALID: ledger x5 for you@company.com (2026-09-20)

• Activate: save the file (e.g. /etc/invar/license.invar) and point INVAR_LICENSE at it, or place it next to your install.
• Renew: replace the file and restart the Ledger.
• Lost it? Email us from your purchase address; we re-issue. Licenses are files, not activations — no server outage can brick your install.

What INVAR proves — and what it doesn't

We'd rather you read this than discover it.

Cross-hardware bit-identical inference is a separate, verification-grade profile from Anomly's exact-arithmetic work (slower, optional). The full threat model, including accepted risks, is in the repository.

FAQ

Is this just another local-LLM runner?

The running is llama.cpp, credited plainly. What INVAR adds is the part nothing else has: a re-executable, hash-chained receipt for every inference, and a team plane that turns those into audit evidence.

What data leaves my machine?

With no Ledger configured: nothing, ever — no telemetry, no account, no phone-home. With your Ledger configured: receipts go only to the server you run, with your token, over your TLS.

Same prompt on another machine gives a different answer — broken?

No. The default profile pins reproducibility to a deployment. Cross-machine bit-identical inference is the exact-arithmetic profile (slower, optional).

Can't I just delete the worldline file?

On your own box, yes — you own your records. In a team, agents push to the Ledger the moment each receipt is created, so the fleet copy survives local deletion, and the Ledger refuses rewritten histories at the door.

GPU support?

INVAR wraps your llama.cpp build — if yours uses your GPU, INVAR does too, receipts included.

CLI reference