Central LLM WikiHelpSign in← Home
Help · 4 / 5

Self-hosting

Run Central LLM Wiki on infrastructure you control

Your deployment

Central LLM Wiki is software for you to run on infrastructure you control. The repository's docker-compose.yml starts the Bun application and PostgreSQL 16, on a plain Docker host behind your own HTTPS reverse proxy or in Coolify. You configure the public APP_URL, the secrets, SMTP for verification and invitation mail, and your own backups; account and wiki data remain in that deployment. The deployment guide covers configuration, upgrades and key rotation step by step.

First Admin and invitations

  1. Configure SMTP, set INITIAL_ADMIN_SETUP_TOKEN to a random value (for example openssl rand -hex 32) and redeploy.
  2. Open https://orakel.marcobartsch.de/setup, enter the token, your name, email and password, and follow the verification link.
  3. Clear the setup token and redeploy. Setup completes once and stays closed afterwards.
  4. Sign in and finish the setup wizard. Then invite people from Settings → Members & roles; every new account, including Google sign-in, needs an invitation for its email address, and invitations expire after 72 hours.

Optionally, an Admin can map the workspace to one Multica workspace under Settings → Data & integrations to import and reconcile its projects. Project sync never changes memberships, credential grants or agent assignments.

Install an agent skill

In the setup wizard or Settings → Connect an agent, download LLM Wiki Reader for read-only questions and research, or LLM Wiki Librarian for ingest, requested edits, maintenance, and tasks that save results. Install the downloaded skill directory using your agent's documented skill flow; Central LLM Wiki does not install it for you. Keep the skill folder intact, including SKILL.md and agents/openai.yaml.

Connect the corresponding central-llm-wiki-read Reader or central-llm-wiki-write Librarian MCP profile separately, with the required credential and grants. A skill does not contain a token or grant access. Set APP_URL to the public application origin before downloading so the package points to this instance's /mcp endpoint.

The web agent is optional. To use it, configure a model provider and API key under Settings → AI provider. The key is encrypted in your database and requests are sent to the provider you select. Your provider's own usage terms and charges apply; Settings and Stats show token counts and estimated model costs for reference.

Version history defaults to 90 days for older snapshots; set VERSION_RETENTION_DAYS to choose a different retention window. The repository ships no backup automation: back up the PostgreSQL volume with your own tooling. The canonical deployment runs one application instance, and production bearer credentials and OAuth require HTTPS.

FAQ

Pasting a URL says it can't extract the article?
The page most likely has a bot check or needs a login; JavaScript-rendered pages are already retried with the built-in browser. If it still fails, copy the text and use Paste text, or upload a PDF. Paper pages with a DOI keep their bibliography even when the body can't be fetched. Login-only platforms such as Facebook can't be fetched; paste the text instead.
The sign-up page says I need an invitation?
Every account needs a valid invitation for its email address, including Google sign-in; there is no open sign-up. Ask an Admin of your instance to invite you from Settings → Members & roles. Invitations expire after 72 hours.
Local auto-ingest failed?
The progress panel shows why. The most common cause is the model being rate-limited upstream; the system retries three times automatically. Models on free shared quotas hit this especially often, and switching to a paid model is much more stable. Pages already created are kept.
Which language does the agent write in?
It follows the workspace language (changeable in Settings); schema rules can override it, e.g. "summaries in English, my comments in Chinese". Content can be in any language.
Will a remote agent and the web fight?
No. When the same page is edited from both sides, the later writer receives the current version and is asked to edit again; the agent knows this rule too.
The agent got it wrong?
See Rules and templates: fix or roll back the page, correct it in chat, or write the rule into schema/.
No API key?
You do not need one if you connect an MCP client such as Cursor or Claude Code: it runs its own agent. A key under Settings → AI provider is only needed for the optional local AI in the browser.
My agent cannot write?
Reader credentials and OAuth connections are read-only by design. Create a separate Librarian credential under Settings → Connect an agent with explicit write grants for exactly the General or project scopes the agent needs.

Getting help

For deployment-specific questions, contact the administrator of your instance. Bugs and feature requests are welcome as GitHub issues. See the source and deployment information on GitHub.