Knos

Two contributors, or two agents, changing the same thing without knowing it. Knos is the record of who is on what, and it refuses to answer about work somebody else has already taken.

Start with the pull request check. One file in your repo, nothing installed, never fails a build:

# .github/workflows/knos-claims.yml
on: [pull_request]
permissions: { contents: read, pull-requests: write }
jobs:
  claims:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - uses: drexthealpha/Knos/action@v0.1.4

It reads .knos/decisions.md — a file a maintainer commits — and comments when a branch touches work somebody has claimed or a decision already recorded. It exits 0 on every path, including every failure path (pytest tests/test_shared_repo.py -k never_returns_non_zero). The tag is pinned rather than a branch, so what runs in your CI is a fixed file you can read: git show v0.1.4:action/knos_pr_check.py.

pull_request_target works too, if you want the check on pull requests from forks: the Action reads only the committed .knos/decisions.md and the pull request's own title, body and file list, and never checks out or runs anything from the head branch.

The whole loop, in under a minute

pip install knos && knos connect     # once, per machine
knos claim "the parser"              # agent A takes it
# ask any other agent about the parser — it is refused, and told who has it
knos done                            # give it back
knos export                          # writes .knos/decisions.md, commit it

Claim, be refused, release, export. The Action then does the same thing on a pull request, for people who have never installed Knos.

All of it runs without an editor open, as two real processes against one store: python scripts/withhold.py. The other half of the proof is python scripts/gate.py, which deletes the store and asserts that both the withholding and the answers die with it — scripts/README.md.

Under it is a local MCP server: three tools over stdio — search, about, remember. No HTTP server, no ports, no account, no model download, no repo to register, and no network connection at all: that last one is a test, not a promise.

knos connect adds Knos to Claude Code, Claude Desktop, Cursor and OpenCode — whichever you have, each in the shape it reads (mcpServers for the first three, mcp with "type": "local" for OpenCode). For Claude Code it runs claude mcp add --scope user, which registers the server with the session you are already in, so its tools work immediately with nothing to restart.

For the other three, knos connect writes the config and takes a backup, and then there is exactly one thing left to do:

Every file it touches is copied to <name>.before-knos first, and knos connect prints the restart line for each client that needs one — quit the app and start it again, because nothing in the MCP spec lets a server register itself with a session that is already running.

/plugin marketplace add drexthealpha/Knos
/plugin install knos@knos

{ "mcpServers": { "knos": { "command": "python", "args": ["-m", "knos.mcp"] } } }

The first thing an agent asks about a repo reads it. Seven cold runs each, whole process, Windows on a spinning disk (WSL on the same box: 3.4s median on a small repo): 1.7s median on a small project (1.5-2.1), 2.0s on goose (1.8-3.4), 3.1s on the Linux kernel (3.0-6.8) — 93,703 tracked files. What it does in that time is one git log, your CLAUDE.md, and the transcripts of past sessions in that tree, written to SQLite. It happens once. Every question after it is under 0.2s, and every agent you have shares the result.

What a claim does here

Claude Code is rewriting the risk guard. You ask Cursor about it.

Every tool in the table below answers, and Cursor gives you a confident plan built on the version that was on disk five minutes ago. Knos does this instead:

Withheld. risk guard (held by Claude Code) is being worked on right now,
so knos is not the place you find out about it. Ask them, or work on
something else.

Not a warning attached to the answer — no answer. Your agent can still take it, by saying why, and the reason is written down under its name where you will read it. knos done releases it, and so does half an hour.

Every cell below comes from that tool's own README or documentation page, so you can check each one:

Knos is not the only tool with claims, and that column would be dishonest if it implied so. Vibsync, CoordMCP, AgentRoom and Agent Claim MCP all let an agent claim something; agentmemory has leases. The difference is what a claim does. Everywhere else it is a signal about a file, which an agent may check before editing and may ignore — Vibsync says so itself, and CoordMCP's locks stop edits while the memory stays fully readable. Knos changes what the memory says: ask about work someone else claimed and there is no answer to act on, and the claim is bound to the connection that made it, so an agent naming itself the holder is still refused.

Two of these are worth your attention for reasons other than that column. Memryzed is local, keyless and one SQLite file — the same shape as Knos, with more recall tools and no coordination. Vibsync is the only one that shares a claim across machines, which Knos does not: a committed .knos/decisions.md is as far as a claim travels here.

Knos is not trying to out-remember these tools. It is trying to be the one that speaks up while two agents are in the same code, and to cost you two commands and three tools to find out.

One agent is enough to see it

knos claim "the parser"

Ask your agent about the parser. It is refused, and tells you so. knos done and it answers again. That is the whole mechanism, in two commands.

flowchart TD
    P["pip install knos<br>knos connect"] --> CC["Claude Code"]
    P --> CU["Cursor"]
    P --> CD["Claude Desktop"]
    P --> OC["OpenCode"]

    CC -->|"first question reads the repo"| S[("one SQLite file<br>no server, no model")]
    CU --> S
    CD --> S
    OC --> S

    R["CLAUDE.md, AGENTS.md, ADRs<br>commits, past sessions"] --> S
    W["every worktree of this repo"] --> S

    S -->|"answers, with the source"| CC
    S -->|"withheld - Claude Code is on it"| CU

    style S fill:#1f2933,stroke:#7b8794,color:#ffffff
    style P fill:#e8f0fe,stroke:#4a6fa5,color:#111111
    style R fill:#f5f5f5,stroke:#999999,color:#111111
    style W fill:#f5f5f5,stroke:#999999,color:#111111

Every arrow above is a command you can run: the withheld one is knos claim "the parser" then asking a second agent, and pytest tests/test_intent.py -k withholds_the_answer is the same thing as a test.

Share it with the repo, not a server

Everything above is local. Two things, though, exist nowhere a teammate can reach — what somebody decided, and what somebody is working on right now. So those go in the repository, as a file you commit:

knos export        # writes .knos/decisions.md
git add .knos && git commit -m "share decisions"

That one file is the whole mechanism. It is markdown, it diffs in review, and three different kinds of reader consume it without installing anything:

• A teammate clones and asks. .knos/decisions.md is one of the decision records Knos already reads, so a clean clone answers from it on its first question — no install, no sync, no account, no server between the two machines (pytest tests/test_shared_repo.py -k second_clean_clone).
• Every agent on their machine reads it too, through the same three MCP tools, on the same first question. A contributor who has never heard of Knos still gets your decisions, because their agent asks and the file is already in the checkout.
• CI reads it on a pull request and says so when the branch touches work somebody has claimed (pytest tests/test_shared_repo.py -k ci_warns).

# .github/workflows/knos-claims.yml
on: pull_request
permissions: { contents: read, pull-requests: write }
jobs:
  claims:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - uses: drexthealpha/Knos/action@v0.1.4

The comment is a heads-up, never a failure — action/knos_pr_check.py exits 0 on every path, including when it finds a conflict and when it crashes.

It runs both ways, and that is the part that compounds. The teammate who cloned runs knos export too. Their decisions and their claims land in the same file, you pull, and your agents read what they decided while you were not looking. Every contributor writes to the file and reads from it, so it holds more after each one than before. The repository is the shared object: there is no database of ours, no protocol to adopt, and no server to pay for.

Secrets do not travel: a note about .env is filtered out of the exported file by the same check that hides .env from an agent (pytest tests/test_shared_repo.py -k private_note).

Nothing outside this repository has adopted it yet. The loop is implemented and tested, six tests in tests/test_shared_repo.py. That is a mechanism that works, not a network that exists.

Who else has written this problem down

Nothing here is about Knos. These are other people's open issues, in their own repositories, filed before they had heard of it.

GitHub issues created since 1 January 2026 that use the phrase "multiple agents" "same file" number 1,556; "two agents" in a title or body returns 31,411. Most of that is unrelated prose. The ten below are ones read individually: every one is open, in a repository between 1,212 and 388,747 stars, and describes two agents working over each other.

Two of those — openai/codex #36719 and #37226 — ask for a record of what each agent intends to edit, so that two of them do not edit the same thing. That is the mechanism in What a claim does here. Both are open.

Check the two counts, and that these are still open, with the GitHub CLI:

gh api -X GET search/issues -f 'q=is:issue "multiple agents" "same file" in:body created:>2026-01-01' --jq .total_count
gh api repos/openai/codex/issues/36719 --jq .state

One repository that coordinates agents through a GitHub issue today

tsz-org/tsz runs its agents against a claim board: issue #17314, 1,133 comments at the time of writing, the successor to #15994, which was forked after it hit the 2,500-comment cap. Agents post CLAIM, DONE, DROP and BLOCK lines and are expected to read the board before starting. Its standing gotchas record the failures that follow: sessions that end on a live claim with no closing record, and two claims naming one defect in different words.

That is the same coordination record Knos keeps, kept by hand in a comment thread, at a scale where reading it before every edit stops being possible.

No repository outside this one has adopted Knos. The issues above are evidence that the problem is real and written down by other people. They are not evidence that anyone has adopted this answer to it.

Check any of it in under a minute

Nothing below is a claim. Each row is a command; run it and see.

Cost: pip install knos. No account, no key, no server, no model download, no network request, and a 5 MB free-tier cap per repo.

Everything else, briefly

Every answer names where it came from — a commit, a session and a date, or a file and a line. Knos has no model: it does not summarise and it does not guess, it finds what somebody actually said.

$ knos ask "what are the rules here?"

Ask before adding a dependency. The build is the product.
    AGENTS.md:3
Every change ships with a test. A green run you did not watch is not green.
    CLAUDE.md:8

What is covered, and what is not

Knos is wired into 4 clients and reads the past session history of 2. Both numbers are the honest ones:

Wiring a client is three edits and a test — CONTRIBUTING.md has them, with OpenCode as the worked example. A session reader is about 40 lines: Codex CLI and Gemini CLI are the two missing ones, each a small parser plus one test, written up in .github/GOOD_FIRST_ISSUES.md.

Knos does not have memory of every local workflow, and does not claim to.

.env, *.pem, id_rsa, .ssh, .aws and twelve more are private the moment Knos reads a repo, without being asked. Private means invisible, not redacted: an agent asking about one is told nothing at all — no result, no count, no "2 hidden".

Worktrees. Keep them; they do a different job, and Knos treats every worktree of a repo as one memory anyway. Read the repo in one tree and every other tree can answer. Claim in one and the agents in the others are held off.

The whole mechanism is one git command. git rev-parse --show-toplevel returns the worktree root, so every worktree looks like a different project; git rev-parse --git-common-dir returns the git directory the worktrees share, which is identical across all of them. Knos keys the store on the second. Tools that key on the first fragment a repo's memory once per worktree — that bug, in another tool. Check it: pytest tests/test_worktrees.py.

A claim lapses after 30 minutes. An agent that crashes mid-change never calls knos done. If the claim outlived the process, that work would be unaskable until a person noticed and cleared it by hand. Instead the hold expires on its own, and the next agent to ask gets a real answer. Taking a claim is a compare-and-swap, not a blind write, so two agents reaching for the same work in the same second do not both believe they have it: one wins, the other is told who holds it. Check both: pytest tests/test_intent.py -k "lapses or two_processes".

Five tiers, one file, a hard 5 MB cap. Sibyl's schema is not a black box Knos writes blobs into — it uses the tiers for what they are. Live claims go in HOT, one row per topic, overwritten rather than appended. Decisions and files go in WARM. History goes in COLD, append-only. The whole thing is capped at 5 MB by Sibyl's free tier, and knos status prints the size and says nearly full from 4 MB, so a store that is filling up tells you before it stops taking writes rather than after (pytest tests/test_sibyl_is_load_bearing.py -k cap_and_warns). At 5 MB a claim is refused in words, not dropped: an agent that thinks it holds work it does not is the exact failure this whole feature exists to prevent (pytest tests/test_sibyl_is_load_bearing.py -k full_store).

Commands. knos ask, knos claim, knos done, knos status, knos export. knos help lists the rest. Nothing runs itself: no watcher, no daemon, no schedule.

Speed, on the one question this is for

"What was decided, and is anyone on it?" — warm, whole process, median of 7:

Git wins on a small repo and it is not close. Knos's time is flat with repo size because it reads an index rather than walking history; git's grows with it. On the kernel that is 30x, and most of Knos's 900ms is Python starting up.

Knos is not faster than git at anything git is for, and a cold first read of a large repo is slower than either — 3.1s median, stated above.

What agents actually read. A study of 557 agent sessions and 33,097 pull requests found that 60.5% of everything coding agents do with documentation happens in instruction files and their own notes — CLAUDE.md, AGENTS.md, plans, scratch notes — against 10.6% for classical docs and 1.3% for API references (Gao & Chen, 2026). The same paper says the link between what agents consult and what they edit is unresolved, so this measures where agents spend their documentation time, not that Knos is needed. What is separately true: none of those files can say who is reading them, or what another agent is changing right now.

What happens when you delete the memory

One SQLite file at ~/.knos/<repo>/memory.db, via Sibyl, capped at 5 MB per repo — Sibyl's free tier, and Knos runs it unactivated, so there is no account to make and no cap to raise. Sessions and commits are read newest first, so when a repo fills, what you have is the recent end of both and the older end was never read. Nothing already stored is evicted or truncated, and knos status says nearly full from 4 MB. The Linux kernel filled 0.3 MB.

Nothing leaves this machine — Knos makes no network request. Delete that file and:

knos status counts that first row for you, so you never take it on trust:

journal    330 things learned
             0 of them exist nowhere else - told, claimed, stood down
             delete the store and only those go; the rest is re-read from your repo

Ten seconds to prove it: claim something, watch an agent be refused, delete the file, ask again. Nothing was ever held.

More on the five tiers, why a claim expires, and how a hold is bound to a connection so an agent cannot borrow somebody else's name: docs/core-flow.md.

The two onchain parts, and exactly what they are

Both are optional. Knos works with neither, and nothing on the read or answer path touches a network — that is what pytest tests/test_no_network.py checks.

Base: sharing one folder with a teammate

knos share ./src --with alice.base.eth
knos unshare ./src --with alice.base.eth

What it does. Their agent can read that folder and nothing else. The record of who may read what is Access.sol on Base Sepolia, so neither machine has to trust the other's copy of the answer. Testnet, so it costs nothing.

What it does not do. It does not move your memory anywhere — the store stays on your disk. It does not encrypt anything. It is one permission bit per person per folder, not a sync protocol.

How to verify it. Two commands and one number each way:

python -c "from knos import team; o=team.identity('owner').address; m=team.identity('teammate').address; \
team.share('crates','teammate'); print(team.may_read(o,'crates',m)); \
team.unshare('crates','teammate'); print(team.may_read(o,'crates',m))"
# True
# False

Or read it without running anything: contract 0x955fa320…6E52, and one full cycle — deploy, grant, revoke. Nine contract tests: cd contracts && forge test.

Virtuals: selling one answer

What it does. Knos is registered on the Virtuals marketplace as a provider with one offering: another agent pays 0.01 USDC for an answer out of this machine's memory. The seller is agent/offering.ts.

agent/bot.ts is that same agent with a chat face, in one process: it answers ACP jobs, it answers /ask out of the same store, and /brief buys something over x402 on Base and writes what it bought back with knos remember, so the next agent on the machine gets it without paying. Every one of those paths reads or writes the same SQLite file — python scripts/gate.py deletes it and none of them work.

Started with no Telegram token it reads commands from the console, so the whole thing runs without an account:

$ npm --prefix agent run bot -- /status
  hot        nothing in progress               one each, expires after 30 min
             0 claims held right now - nothing is being withheld
             2.2 MB of 5 MB used

The x402 half is live on Base mainnet. /brief BTC pays 0.01 USDC to x402-seller's market-regime endpoint and writes what it bought into the store with its receipt. Five settled so far, signed by 0xEca35a0C…48C1: 0x2ce6af5c…, 0x20983f7b…. The client is src/knos/buy402.py, which signs with the keystore knos made itself — there is no private key in any config file.

Two routes on that seller, /markets and /signal, return 502 after the 402. They cost nothing (the payment never settles) but they are why /brief is the only route wired in.

What it does not do. There is no evaluator and no reputation system, and every job traded through it was bought by a test agent of mine, not by a customer. It is off by default and runs only when you start it.

How to verify it. The agent page is public — open app.virtuals.io/acp/agents/01a05b97… and you will see the registration without installing anything. Job 75659 is on Base mainnet, in two legs, neither of which needs an account to read: buyer pays 0.01 USDC into escrow, then escrow releases 0.0095 to the provider — the missing 5% is the protocol's fee.

It was asked why does knos withhold claimed work, and what it sold, in 180ms, was a passage out of a session from four days earlier:

knos withholds what it knows. A second agent searching claimed work gets who holds it and nothing else — the content is absent from the reply, not annotated. — Claude Code session 4101eeab 2026-08-31

Nobody re-typed that. Another agent paid a penny and a fresh process read it back with its source. The buyer was knos-buyer, an agent of mine registered to prove the path executes. It is not demand.

What breaks without the store

Every capability below reads or writes the one SQLite file. The middle column is the command that exercises it; the right column is where it touches the store. python scripts/gate.py deletes the file and asserts the first three rows stop working.

The claim lives in HOT because it is about now and is overwritten, not appended. Decisions live in WARM because they are named things replaced in place. History lives in COLD because it is append-only. That is Sibyl's schema used as intended rather than as a key-value bucket, and knos status prints the tiers by name.

Tests

pytest runs the critical path only — claim, withhold, concurrency, no-network, three tools, private files — 11 tests in about 25 seconds on an idle machine, because a suite you wait four minutes for is one you stop running. The whole suite is pytest -m "": 214 tests, about four minutes. The contract has 9 more: cd contracts && forge test.

Including the ones that would catch a lie:

• test_no_network.py breaks socket.connect, bind and getaddrinfo, then reads a repo, answers questions, writes, claims, withholds and overrides. Nothing reaches for the network, and the guard itself is tested so the test cannot pass by doing nothing.
• test_private.py asks the search layer directly, with an agent's identity, for a private path. Nothing comes back.
• test_memory.py has a second process write a conflict, rejected by the schema rather than by Knos.
• test_recall.py writes as one agent and recalls in a separate, fresh process.

What it cannot do

• A claim withholds what Knos knows. It cannot stop an agent editing a file — nothing on your machine can, short of file permissions. If you need that, use a worktree.
• Rules are enforced only on what Knos mediates: recall, remember, and the claim/withhold path. Knos cannot make a foreign runtime obey your CLAUDE.md; it can only decline to be the source of truth, and say who to ask. No MCP server can do more than this: the protocol gives a server no way to intercept an edit.
• Claude Code and Cursor only. No Gemini CLI or Codex history yet.
• It does not write the answer for you, and it does not watch files. Run knos point again to catch up.
• Retrieval is lexical, not semantic. Sibyl searches with SQLite FTS5, so Knos finds passages containing your words and ranks those. Ask about something the sessions never discussed and you get confident, well-sourced passages that share a word with your question and nothing else: "why did we drop redis" matches every note about dropping something. Ask in the words the work was done in and it is sharp. There are no embeddings at any Sibyl tier — the paid tier adds summarising and a learning loop, not search.
• 5 MB per repo.
• Four jobs have been traded through the Virtuals provider, all bought by a test agent of mine. Nobody else has bought anything.

Contributing

CONTRIBUTING.md has the three edits an agent adapter takes and the test to copy. pytest runs the critical path in about 25 seconds; pytest -m "" runs all of it in about four minutes, against throwaway stores.

Licence

MIT.