wikai
Download

The chat sidebar

The chat sidebar is a panel on the right where you ask questions about the page you’re reading or the whole wiki. Answers are grounded in your own wiki, and the agent’s edits appear live in the main view.

Connect an agent

Wikai is bring-your-own-agent: it doesn’t ship its own model, and AI usage is billed by your provider, not by Wikai (no token markup). Six agents are available, each on its own tab under Settings → Agents, and every one of them works the wiki through the same tools with the same rules — an edit to a protected page arrives as a suggestion for you to approve, whichever agent proposed it.

If you have a Claude Code subscription, Wikai uses your existing Claude Code install — there’s no API key to manage. Wikai finds the claude binary automatically.

Authenticate once per machine by running this in a terminal:

claude /login

Wikai never sees your subscription token — it uses the credentials Claude Code stores in ~/.claude/. If Settings shows “Found, but not signed in”, run the command above and re-detect.

Claude Code reaches your wiki through Wikai’s own MCP server. If it starts a turn without that connection — the wikai server is disabled for the project your wiki’s folder belongs to, say, or it failed to connect — the sidebar says so in an amber line at the top of the turn, with the reason, rather than letting you find out when every wiki tool fails one call later.

Codex

If you have a ChatGPT subscription, Wikai runs the Codex CLI you already have, the same way it runs Claude Code: the binary is found automatically, there is no key, and you sign in once with codex login in a terminal. Codex runs with its own file access switched off, so — like every other agent — it can only change your wiki through Wikai’s tools.

OpenRouter

One API key gives you a long list of hosted models. Paste the key, and Wikai fetches the models that can call tools and offers those; pick one in the model dropdown. Your prompts and wiki context go to OpenRouter and on to the model’s provider. OpenRouter is metered, so daily-note enrichment won’t run automatically on it.

llama.cpp and vLLM

A model you run yourself — on this machine, on another one you own, or on a server you rent. Start llama-server or vLLM with an OpenAI-compatible endpoint, paste its address, and Wikai probes it and tells you what it is serving. The Settings tab says in one sentence where your pages are going, because that depends on the address you typed; a server on the open internet with no key gets a warning. A local model is free to run, so automatic enrichment is allowed. These two are marked beta: they work, and they have been exercised less than the rest.

Anthropic API

Paste an Anthropic API key and pick a model. Useful if you have a key and no Claude Code subscription.

Coming soon

A Grok tab is already in Settings so you can see where it will live; it does nothing yet.

Choosing the model

Next to the agent picker in the sidebar header is a model dropdown. It lists what the selected agent offers: Codex reports its catalogue, OpenRouter and the local servers are asked for theirs, and Claude Code takes an alias — opus, sonnet, haiku, fable — for the latest model of that family. Default means the agent picks, which is the one choice guaranteed to be inside your plan. Wikai remembers the model per agent, and after the first turn the dropdown’s tooltip shows which exact model Claude Code resolved an alias to.

Codex adds a reasoning effort dropdown. The options come from the model you chose — they differ per model, and a wrong one would cost a failed turn, so you can only pick a right one.

A reopened chat goes back to the agent it ran on, and a new chat starts on your default agent — or on the one you used last, if you set the default to Last used.

Grounding

The agent answers from your sources, not the open web (unless you turn on internet search). The whole wiki is always its context, and with every message it is told which page you have open — so “summarise this page” or “what links here?” work without naming the page, and a question about something else isn’t pulled towards whatever happens to be on screen. There is no mode to set; before 0.5.2 a Ground in selector asked you to choose.

Your comments on a chapter are part of that context too: the agent can read what you wrote in the margin of a page and add, change or remove notes of its own there, under the same edit policy as any page.

What the agent can do

  • Answer questions with citations into your wiki.
  • Draft new pages and edit existing ones — edits respect each page’s agent_edit policy, so changes to protected pages arrive as suggestions you approve rather than silent rewrites.
  • Propose a research project when a question is bigger than a chat — but only propose. You get a banner, and no work document is created until you accept. Asked a plain question, the agent answers it.
  • Tool calls show up in the transcript as it works; a Stop button interrupts a turn.

Last updated 2026-10-07.