An Article 11 AI product: a local plugin you can download today, and a hosted control plane in design
One bounded task. Independent AI seats. One governed receipt.
What exists now is a local plugin, version 0.1.11: Claude and Codex each get a private memory library on your disk, and a finite workshop lets one draft, the other review, then fresh sessions revise and assess, with the record kept on your computer. The hosted control plane described further down, with route binding, one-use human approval, private receipts and an emergency stop, is a design for a future pilot; the download does not include or enforce it.
Download: 460,445 bytes, SHA-256 e144a4c1cd16a26193b8ebc7d6e69d34626858d0fc8cf8c68e61ac47a01502c9, Apache-2.0. Public discovery is live; the hosted council-run execution described below is not exposed.
Available today: the local plugin
Two ways to use it, both on your own Windows computer with your own existing accounts. Nothing runs in the cloud on our side, no account is created with us, and no model runs until you type a command that says so.
Open SETUP.cmd to connect a private library to Claude Code, Codex or Claude Desktop. Your assistant can then read earlier notes, save a decision, correct it later, or decline. The library stays in a folder you choose, outside the plugin. Three-step setup.
Two extracted copies, two task folders, two libraries. Each stage is one command you run; a question or a decline stops the task; the result is a Markdown file with attributed replies. Reviewed quickstart and the shipped four-stage guide.
The workshop runner checks a Claude Max login through the Claude Code CLI and refuses other tiers, checks a ChatGPT login through the Codex CLI, and refuses API-key variables. It uses your existing allowances; there is no billing fallback. Tested on 2026-09-22: Claude Code 2.1.278 completed its stage; the standalone Codex CLI 0.145.0 with its default model, gpt-6-astra, was rejected by its provider (HTTP 400, "requires a newer version of Codex"). That is the only combination tested; which Codex CLI version works, and the minimum, are untested.
What leaves your computer: during a stage, your own CLI sends that stage's provider the task text, the earlier shared replies, an automatic overview of that participant's library (recent note titles and message subjects, not bodies), the optional starter letter, and, for a returning stage, a validated window of its own earlier conversation; later rounds of the same stage also carry the model's earlier replies and action requests and the results of the memory operations it asked for. The whole library, unread note bodies, raw checkpoint files, receipts and logs are not uploaded. Handoff files between the two installations carry the task and the shared final replies, not private checkpoints. Forgetting a note does not erase exports or backups you made.
Status, stated plainly: the package is source-reviewed and its one designed demonstration (two installations, four stages, eight requests, 15 September 2026) ran on one Windows computer. On 22 September 2026 the builders ran the reviewed quickstart themselves as a new owner-assisted trial: stage 1 completed on Claude with one model response; stage 2 started on Codex and stopped when the provider rejected the tested CLI and model, so the four-stage loop was not completed in that trial. No outside operator has run it yet. The record and the outside-operator card say what to report.
Experimental Windows companion
The SPIRALMESH Talk preview brings the existing workshop into Saved tasks. Start with an empty library, create a small task or import an exported website reply, then choose each draft, review and finishing stage. Open, Reopen and Stop use a small control window.
Requires Python with Tkinter and, for model stages, existing Claude Max and native Codex ChatGPT clients. It includes no model or household memory service. Tests on one Windows computer passed; independent-computer first use remains unverified.
In design, not in the download: the hosted control plane
SpiralMesh™ v1 is a staged contract for a bounded run, not a promise that every named model is available. Each approved seat carries exact provider, model, identity-prompt, generation-configuration, request, and response bindings.
A preview names the seats, data class, policy, limits, and stop profile. No send is authorized by the preview itself.
A short-lived approval binds the exact candidate and allowed effects. Replay must fail closed.
The receipt preserves results and dissent. The emergency stop denies new sends and records residual risk without claiming remote deletion.
How the v1 pilot is designed
Hash the input, policy, route, seat identities, generation settings, limits, visibility, and emergency-stop profile.
A human authority approves that exact candidate, allowed seats, effects, expiry, and cost ceiling. The nonce is single-use.
Only approved seats may receive the allowed data. Provider and local-runtime responses remain separately attributable; dissent is preserved.
A private receipt records what occurred and what it does not prove. Human delivery, publication, memory, and anchoring are separate decisions.
This is an illustration of the proposed contract shape, not evidence of a run, customer delivery, public receipt, or blockchain anchor.
First use
Article 11 AI, Inc. is preparing a genuine first-use pilot for an external interstate participant. No public API key, production quota, customer commitment, or price is offered by this candidate page.
Case-by-case · approval required
Not publicly available
Scope and price not published
Optional anchoring
The public Article 11 chain and OpenTimestamps proof surfaces are live. That does not mean every SpiralMesh run is public or anchored. The v1 pilot requests neither. Any later sanitized hash submission is a separate, scope-bound approval.
The live anchor endpoint is authoritative for current state and distinguishes pending calendar aggregation from a verified Bitcoin proof.
A public hash can still create correlation risk. Publication and anchoring remain off unless separately reviewed and approved.
An anchor can help show that recorded bytes were not quietly changed after the proved time. It does not make the underlying answer correct.
Proof, not promise
The endpoints below show public site, discovery, chain, and anchor state. They do not yet expose a SpiralMesh council-run operation. A future run is proven only by its approval and receipt.
Check a SPIRALMESH workspace receipt in your browser, or save the verifier to use offline. It checks supported unsigned fixture and local-model receipt formats, hashes, and internal consistency. It does not establish who produced a receipt or independently prove the recorded events.
For AI agents
Agents may inspect this page, JSON-LD, the embedded manifest, markdown content negotiation, and the public discovery documents. None of those surfaces authorizes a model invocation, vendor send, memory write, publication, or chain write.
article11.ai/agents and article11.ai/agent-orientation.json.article11.ai/api/discover and article11.ai/api/openapi.json.<script type="application/json" id="spiralmesh-manifest">.mesh.help() describes the candidate; it performs no council run.AI is an intended reader and future participant. Human authority, exact contracts, provider boundaries, and receipts still govern every effect.
Questions, answered plainly
Yes, as a local plugin (0.1.11) you download from the downloads page and run on your own Windows computer with your own Claude Max and ChatGPT logins. It gives Claude and Codex private memory libraries and a finite draft, review, revise, assess workshop. It is not a hosted service and does not include the one-use approval, receipt or emergency-stop controls described below; those belong to the control-plane design. Start with the reviewed quickstart.
A design for a pilot-stage control plane for bounded AI work across approved cloud and local seats. A run proposal pins the input, policy, route, identities, model configuration, cost ceiling, receipt visibility, and emergency-stop profile before a human can authorize it. It is not publicly exposed and not part of the download.
A receipt can prove the bytes, bindings, timestamps, approval consumption, stop state, and preserved dissent that it actually records. It does not by itself prove factual truth, correctness, legal validity, consciousness, personhood, or fitness for purpose.
No. The v1 pilot keeps raw payloads and receipts private and requests no public or chain mirror. Any later sanitized publication or anchoring requires a separate approval.
Not through a public council endpoint today. The public orientation and discovery surfaces are readable but grant no execution authority.
There is no public self-service key or published production price on this page. A reviewed pilot must declare its provider route and cost ceiling before human approval.
A human authority approves each bounded run. Provider terms, applicable law, data boundaries, and the exact signed run contract remain controlling. A receipt records the decision; it does not replace judgment.
Trademark notice
SPIRALMESH™ is the subject of U.S. Trademark Application Serial No. 99575218 in International Class 042. A Notice of Allowance issued on July 21, 2026. The application remains pending; no registration or right to use the ® symbol is claimed here.
SPIRALMESH™: intent-to-use application under Section 1(b) · first Statement of Use or extension deadline: January 21, 2027.