Reconcile the claims

The register

• “Star us on GitHub at launch” / “repo dropping soon”

  Corrected

  The repo has been public since 1 September 2026, under MIT. We were telling people to wait for something that had already shipped. Both lines are gone and the button now points at the repo.

  gh repo view vibemind-space/vibemind-os · 2026-09-24

• “The agentic engine, the OpenClaw desktop driver and the orchestration framework are on GitHub under MIT.”

  Narrowed

  vibemind-os is public and MIT. Vibemind_V1 is public but its default branch carries no licence file, which means it is not open source in any usable sense until one lands. The OpenClaw and OpenFang forks are public, but as forks of upstream projects rather than components we wrote. We name the one repo we can stand behind.

  gh repo view, both repos · 2026-09-24; forks re-checked 2026-10-27

• “MIT licensed”, as a single label for the whole stack.

  Corrected

  The stack is assembled from components under several licences. The execution runtime is Apache-2.0 or MIT — our first correction said Apache-2.0 only, which was wrong too. One brain component is licensed for non-commercial use only, a few components are AGPL-3.0, and more than three carry no licence file, which grants you no rights beyond looking at them. We state the licence per component — see the licensing page.

  gh repo view across all submodules · 2026-09-24; licence files per pinned commit · 2026-10-27

• “Self-host it. Fork it. Audit it.”

  Narrowed

  We first wrote that operations and the security suite were private. That was wrong: both have been ordinary public folders since May. What does block an anonymous recursive clone of the public repository is one private video submodule. Fork and audit work; self-hosting works if you are ready to read configuration, and is not a one-command install.

  gh repo view across all submodules · 2026-09-24; anonymous fetch of every pinned commit · 2026-10-27

• The desktop gateway presented as our own open-source component.

  Narrowed

  It is a fork of a large upstream project, not our own component. When we first checked, GitHub did not classify the upstream licence; it now detects it as MIT, so our MIT label on the fork is consistent with upstream. What we no longer say is that we built it.

  Upstream repository metadata · 2026-09-24; re-checked 2026-10-27

• “Ideas — 163 tools — restructure thoughts into kanban, SWOT, or mindmap.”

  Corrected

  The 11-format dispatcher was deleted on 16 June 2026 and our own audit found no replacement. What survives is an MCP server of about 830 lines doing bubble create/read/update/delete/promote plus idea linking. The formatting feature does not exist today.

  Vibemind audit note, 2026-09-22 · commit ec98a0c6

• “476 specialist tools.”

  Corrected

  That total was counted before the June refactor. We removed it and recounted from the code on 27 October: 50 in-house tool servers with 798 tools between them. That is a count of code, not of what the runtime exposes; the Ideas space, once claimed at 163 tools, has 13.

  Audit note 2026-09-22 · tool count from the code · 2026-10-27

• “Coding — describe an app, get a project the engine builds, verifies and self-heals.”

  Unverified

  The routing is real: the registry holds 13 coding capabilities that post jobs over HTTP to the engine; the two for code preview are switched off because the engine no longer serves that route. The engine’s pinned version can be fetched publicly now; it could not during the September audit. As of 27 October the engine is not running anywhere in our setup, so treat the end-to-end claim as untested until we publish a run you can reproduce.

  space_agent_registry.yml @ 2026-09-08 · audit note 2026-09-22 · live check 2026-10-27

• “EN / DE, sub-second voice”, presented as a headline capability.

  Corrected

  The dedicated voice agent file was removed in the same June commit and the audit could not find it under any path. Voice work is in progress, not shipped. It has moved to the in-progress column where it belongs.

  Audit note, 2026-09-22

• “Thousands waiting.”

  Corrected

  Our Discord had roughly 70 members when the documentation was last written. We do not publish a waitlist number, so we should not have implied one. The line is gone.

  readme_doc.pdf · 2026-09-22

• “MVP shipping today.”

  Corrected

  Our own internal status in August read “all in proof-of-concept state”, and core routing work was still open in September. A shipping date that never resolves is worse than no date. Removed.

  Internal status, 2026-08-09

• “15 People. One Mission.” — with titles for everyone listed.

  Narrowed

  The list mixed full-time builders, part-time contributors and people who had agreed in principle but not started. Everyone on it is real and every one of them has helped. Flattening that into fifteen staff with job titles overstated it. We said in September that the page separated core team from contributors; it did not until 28 October. On 28 October it listed three co-founders, three leads and seven team members, thirteen people in all, each confirmed by the founders; two names were removed. On 30 October the founders added five people and sorted everyone into teams: three co-founders, three leads, and Vibe Coding, Homelab & Infrastructure, Sales, Business & Finance and Ethics — eighteen people — plus one partner with his own product, who is not counted as team.

  Internal review, 2026-09-24; founders’ confirmation, 2026-10-28 and 2026-10-30

• “3 Cities · DE / FR / TR” alongside a footer naming six countries.

  Corrected

  Two different numbers for the same fact on the same page. The footer counts everyone who has contributed from anywhere; the stat counted where the core team sits. The cities stat is gone: the team section now names where each person sits, and the footer names Germany and India.

  Internal review, 2026-09-24; founders’ confirmation, 2026-10-28

• Ethical guidelines and a non-profit structure, with a named owner.

  Narrowed

  This is a genuine intention with real thinking behind it, and no legal structure, no published framework and no board. Our own document labels itself a vision document with no implementation. It is now described as a commitment rather than a thing that exists.

  readme_purpose.pdf · 2026-09-22

• “Where It Actually Stands” — running / in progress / next.

  Corrected

  In September we judged this section honest and kept it. By 27 October two of its “running today” lines no longer held: the central Brain service is scaled to zero and the Bridge module has no caller, so “Brain → Bridge → OpenFang, end to end” was not true; and the coding engine runs nowhere. The column now lists what we verified: the spaces that run on our own server (five since the re-grading of 11 November), the agent runtime, the agent task chain and capability routing in reproducible runs. Video production tooling is real and in use, with its GPU side on a workstation.

  readme_marketing.pdf · 2026-09-22; live check and end-to-end runs · 2026-10-27

• “Regex-matched against a 66-capability registry.”

  Corrected

  The registry holds 119 capabilities, 109 of them enabled. Regex first is accurate and was verified end to end against the live database. Two related claims did not hold: the router does not filter by the active space, and the semantic fallback is broken in the code.

  brain/the_brain/data/capabilities.yaml · capability routing run · 2026-10-27

• “The Bridge dispatches … over one of eight transports.”

  Corrected

  The bridge module exists but nothing calls it. Dispatch happens inside the Brain, which turns a plan into database operations, service calls or tasks for agents. The eight transports are not in use, so the claim is gone.

  Architecture review · 2026-10-27

• “A Rust runtime runs the work inside a metered WASM sandbox.”

  Narrowed

  The sandbox with fuel limits and a watchdog is in the OpenFang code. In operation, none of the 43 agents runs as a WASM module; most run as Claude Code subprocesses with scoped tools. We describe the runtime, not a sandbox it does not use.

  OpenFang agent list · architecture review · 2026-10-27

• “28 providers, 130+ models.”

  Corrected

  Out of date and too low: the runtime has adapters for 42 providers and lists over 200 models. In practice most agents use one hosted Claude model. We now say both.

  OpenFang provider registry · 2026-10-27

• “Eight cognitive collections … 1,400+ plans recorded.”

  Corrected

  There are 13 memory collections. The collection of routing decisions holds 171 entries. We could not find where 1,400 came from, so the number is gone.

  Qdrant collection counts · 2026-10-27

• “56k+ code chunks … self-heals on every commit — a full re-index in about five minutes on GPU.”

  Corrected

  The last full index we can verify had about 67,000 chunks, in August. The automatic updater has not run successfully since 8 September, and the one measured GPU re-index took about 25 minutes. The page now says so.

  fungus-search logs · 2026-10-27

• “40+ channels. OpenClaw receives, OpenFang sends.”

  Corrected

  The runtime ships 42 channel adapters, but two are configured: Telegram, which OpenFang both receives and answers, and WhatsApp, which OpenClaw both receives and answers. Telegram currently goes to an operations agent rather than through the Brain. That Telegram path was run end to end on 27 October: a test message from a user account reached the agent, and its reply arrived in the chat in under 30 seconds. WhatsApp has no published run yet.

  OpenFang channel config · architecture review · Telegram end-to-end run · 2026-10-27

• “13 Spaces. One Brain.”

  Narrowed

  The code has 21 space folders and the routing registry knows 13; neither number says what runs. The spaces section now lists 18 spaces with a status each: five live on our own server, eleven in progress and two research. Marketing and Video are live with generation still on a workstation. Earlier counts on this page were higher because they included spaces that are deployed but not finished (Learning, Project Intake) or that were folded into another (Automation into Captain Cook). Personal Space and Agent Farm joined the list on 11 November; their code is not in the public repository yet. The Brain does not route to several of the spaces that do run.

  Space registry and live check · 2026-10-27; founders’ confirmation · 2026-10-28 and 2026-11-11; vibemind-os spaces/agentfarm/runtime-manifest-v1.json

• “Every agent is open source.”

  Corrected

  Not true while several components carry no licence file and one video component is private. The line is gone.

  Licence files per pinned commit · 2026-10-27

• “A full Mailcow stack sends our campaigns on our own domain … No US email cloud.”

  Corrected

  Mail for our domain runs through a hosted provider, as the MX and SPF records show. The Mailcow path in the marketing code is a leftover. We now say that mail and model calls still use outside providers.

  DNS records and marketing code · 2026-10-27

• “No telemetry.”

  Narrowed

  True for the runtime: it sends nothing home. Not true for this website, which counts visits with an analytics service. Until October it also loaded its fonts from a third party; they now come from the site itself. We say “no telemetry in the runtime”.

  Runtime code · website network requests · 2026-10-27 and 2026-11-11