The neutral meter for AI agent work: neither side keeps the count.

Of first agent pull requests that claimed passing tests, 17.8% had a failed check (147 of 826 repositories).

Try the demo · Check an invoice (shadow mode) · Install

Prices

The rule: Knos never charges the party being rated. Nobody has bought anything. docs/MARKET.md has the fee by order size, one customer worked through and the first steps; docs/PILOT.md is the Pilot in full.

What is real today

Devnet is test mode: everything runs on Solana devnet, and the money is test USDC. Mainnet is not touched.

• What works, and how far each part has got: docs/CAPABILITIES.md, one row per capability with its evidence.
• Who uses it: docs/submission/NUMBERS.md. Nobody outside Knos has funded an order, bought anything or signed anything.
• Who can change it: one person, through a multisig, after a public 48-hour delay (docs/GOVERNANCE.md). No outside security firm has examined anything.

This page names no time for an upgrade, so that it is true before and after one. Whether a proposal is pending, has executed or was withdrawn is read from the chain by knos status, the site's banner and the site's upgrade record (web/upgrades.json is the committed copy). The first deployment (programs) has no upgrade authority and stays as a record; nothing new is funded there.

Reproduced by someone else: none recorded yet. Exercised on devnet: none recorded yet. Deployed on devnet: verify_github, verify_gitlab, key_guardian, fund_by_comment, fund_from_wallet, pay_on_merge, hold_and_bind, refund, pause, meter_single, passkey_payee_wallet, upgrade_gate. Everything else is tested locally, implemented or not built: the table with the evidence has one row for each capability, from docs/capabilities.json. Deployed and exercised are counted only at the public program ids; what the 0.3.14 rehearsal ran at staging addresses of its own is in the note of each capability it ran, with its transaction.

How it works

A work order is a task, its budget and the terms that decide whether it is done, fixed before the work starts. A bounty on an issue is the smallest one.

1. Fund with terms. A maintainer comments on an issue: /knos fund 20 checks: test. The terms are hashed into the order. Nothing changes them afterwards.
2. Accept. Someone opens a pull request, a person or a coding agent, and she merges it. Had test failed at that commit, the merge would pay nothing, and a comment would name the check.
3. Pay, or only count. A workflow at a pinned commit reads GitHub's record of the merge, and GitHub signs the run. A Solana program checks that signature and pays the author. A buyer who pays by invoice uses the count alone: knos-meter records each signed evaluation once, and moves no money.

The seller can settle without the buyer, after a merge in a public repository: knos settle --neutral <pull request URL>. With no payment by the deadline, the amount and the fee go back, and that needs no token.

A check can also pay without a merge, when it is black-box: of 63 cheating pull requests, 56 passed plain CI and 0 passed it (docs/TAMPER.md).

Every command and term (checks, paths, days, warranty, holdback, arbiter, /knos offer, /knos split): /knos help, and docs/SECURITY.md, section 3. Why a signed run and not a description: docs/WHY.md. What else exists, and where it is ahead: docs/COMPARE.md.

Limits

docs/SECURITY.md has all of them. The ones to know first:

• Devnet, test USDC, no outside review. knos mainnet-check fails on that line on purpose.
• A signature authenticates a statement, not the truth. GitHub signs which workflow ran, at which commit, in which repository. It does not sign what the workflow read.
• The seller's own settlement covers public repositories only.
• A passed check is not good work. An order buys what its terms say, and the funder chose the terms.
• Knos inherits GitHub's failures. If GitHub signs something false or is down, the programs believe it or wait.
• The payout is USDC on Solana. No card, no bank transfer, no bank payout.

Documents

• Try it: SHADOW (check an invoice, no install), PLAYGROUND (fund a test order with one comment), INSTALL, CONSOLE, PILOT, the recording (a release with no recording has no such file).
• For finance: FINANCE (the four records and the exports), METER, RECEIPT, TERMS (terms cited by hash), OUTCOMES.
• The verifier: VERIFIER (any RS256 issuer), OIDC, COMPOSE, CONFORMANCE, INTEGRATIONS, X402, ADAPTERS.
• Trust: SECURITY, ASSURANCE, INVARIANTS, PROVENANCE (source to build to chain), UNWRAPS, TAMPER, DRILLS, GOVERNANCE, CONTROLS, PRIVACY, REGULATION, REPRODUCE.
• Numbers: BENCH (every measured number), INDEX (the Agent PR Index by week), LOAD, NUMBERS (outside use, zeros included).
• The business: MARKET, WHY, COMPARE, TEAM, DISCLOSURE.
• Operating it: RELAY, OPERATIONS, RELEASE, CHANGELOG.

What is in this repository

History

Knos 0.1 (1–7 Sep 2026) was shared memory for coding agents built on Sibyl. 0.2 and 0.3.0–0.3.9 (29 Sep – 2 Oct) tried coordination, budgets and a jobs market before the measurement above showed where the problem was (drexthealpha/knos-labs). 0.3.10 to 0.3.15 are the two deployments, work orders, the count and the console. 0.3.16 is this release. docs/DISCLOSURE.md says what was built when and what came from elsewhere.

MIT, all of it. Built by drexthealpha. Its memory engine is Sibyl.