About AGNT

The operating system for accountable AI work

AGNT exists because AI work should be verifiable. We build a local-first runtime where agents hold real jobs, use real tools, and produce receipts for everything they do.

What we believe

Local-first by conviction

Your agents, keys, memory and data live on machines you control.

Verification over vibes

Receipts, approvals and audit trails are the product, not an enterprise add-on.

Open and inspectable

Open source on GitHub — read the code your operations run on.

What we actually ship

How to check any of this for yourself

Every claim on this site should be verifiable by someone who does not trust us, which means telling you where to look rather than asking you to take our word for it.

Setting it up
  1. Run it. Community Core is a free download for Windows, macOS and Linux. It is a desktop application, not a trial account, and it works with no server at all.
  2. Read the code. AGNT is on GitHub. The runtime your operations execute on is the runtime you can read.
  3. Read the API. The local API is documented endpoint by endpoint, including the ones this site’s own pages are generated from.
  4. Check the research. Our whitepapers publish their artifacts — scripts and certificates that replay — so results can be reproduced rather than believed.
  5. Check the comparisons. Where a competitor is better we say which properties and why, and we date the review so you know how stale it is.
What a receipt is for
Illustrative example — not a screenshot
A run does not just report success. It records what was
done, what it cost, and what it decided not to do:

  {
    "status": "completed",
    "initialPrompt": "Send the weekly digest to the team.",
    "provider": "...",
    "totalTokens": 66319,
    "estimatedCost": 0.0906,
    "toolExecutions": [
      {
        "tool_name": "execute_shell_command",
        "status": "completed",
        "input":  { "command": "..." },
        "output": { "success": true },
        "error": null
      }
    ]
  }

The runs worth keeping are often the ones that did nothing:
a digest that was not sent because the source data had not
changed is a decision, and it is on the record.
Limitations worth knowing
  • Local-first is not the same as absolute privacy. Your files, credentials, memory and receipts stay on your machine — but if you point a step at a hosted model provider, that step’s content goes to that provider. Use a local model where that matters.
  • A receipt records what the runtime observed. It is an honest account of the steps taken, their inputs, outputs and cost. It is not an independent audit of whether the work was the right work.
  • The marketplace is largely user-submitted and changes without notice. Read what you install; the source of any installed plugin is retrievable.
  • We are younger than the tools we are compared against. Our node catalog is smaller than n8n’s, our community node ecosystem is newer, and multi-user and horizontal-scale stories are not our design target. Where that makes another tool the better choice, our comparison pages say so.
  • Dated claims go stale. Comparisons carry a review date because both we and everyone we compare against ship frequently.

Next: download Community Core, read the Getting Started guide, browse the whitepapers, or see the security model.

Give AI a job. Get the proof.