Docs ​

Every page of the OpenClaw for Open WebUI docs, with one line on what it covers. Each page opens with a one-line summary, then an Applies to: line (Pipe, Action, Filter) and the valves it uses, so a person or an agent can tell from the top whether it is the right page.

The three functions: the Pipe is the chat integration, the Status Action adds the Status button, and the thinking filter adds the per-chat thinking control.

Install ​

• Quick start: the shortest path from nothing to the first reply, with a check at each step.
• The installer: install.py, its prerequisites, the wizard, environment variables and flags, the no-clone form, what install does, and its commands.
• Manual install: paste the functions in the admin panel and set the valves yourself, with no admin password given to a script.
• Device pairing: approving the device on the Gateway, how the identity persists, STATE_DIR and its /tmp fallback.
• Status button: installing the Status Action, the composer Status button and the rate-limit plugin.
• Thinking control: installing the thinking filter and attaching it to models.
• Agent skills: teaching your agent the two conventions (questions and media), installing and updating the skills.
• Update and uninstall: updating, backups/, rolling back, uninstalling, revoking the device.
• Two hosts: Open WebUI and OpenClaw on different machines: what runs where, the network, approval.

Use ​

• Why this exists: the native protocol versus a text API, and the full comparison with OpenClaw's /v1 endpoint.
• Features: the full feature list, each line linking to its page.
  • Chat: streaming, tool cards, tool failures, image input, media, the status line, auto-title, embedding your own page in a reply.
  • Questions: question cards and the secret dialog, which kind you get, what happens on a reconnect, secrets the agent requests, and plugin approvals asked in the chat.
  • Status: the Status dialog, Compact, the sub-agent list and drawer, the composer button.
  • Sessions and models: one session per chat, the model picker, its size cap, DEFAULT_MODEL.
  • Replies from the agent: replies with no message from you: cron jobs, sessions_send, sub-agent results.
  • Reliability: the connection, queueing, reload and restart safety, recovery, Stop and Send now, redeploys, background tasks, errors.
  • Channels: a mention in an Open WebUI channel opens a thread, one session for everyone in it, and the agent follows it.
  • Owner users: OWNER_USERS, for an Open WebUI that people other than you sign in to.
• Configuration: every valve of the three functions, generated from the code.
• Metadata contract: the Sender and Conversation info blocks every message carries, their stability guarantees, and how an agent should read them.

Fix ​

• Troubleshooting: a symptom table, the errors you can see in the chat, where the logs are, and what to put in a bug report.
• Versions: tested versions, minimums, and what an older release loses.
• Security: the secrets, the media file server, the agent's privileges, owner users, the scripts in your browser.

Contribute ​

• Architecture: how a message travels from Open WebUI to the agent and back, why the Pipe is an async generator, and the repo's file layout.
• Development: editing the src/ fragments, the build and its drift guard, running the tests, driving the integration over the Open WebUI API.
• Rehydration and persistence: how Open WebUI keeps an assistant message mid-run. Read it before touching streaming or persistence.

Design ​

• Agent turns: the design behind replies the agent starts on its own, and how each run maps to one message.
• Ask-user modal: a spec for a tool-call-based question flow. It was not implemented as written: the shipped path intercepts OpenClaw needs input: text.