Wikai has always been bring-your-own-agent. This release makes that literal:
six agents to choose from, the model picked beside the chat, and a daily note
that understands time.
Six agents
Until now the chat sidebar spoke to Claude Code or, failing that, the
Anthropic API. It now speaks to six, and every one of them reaches the same
wiki tools with the same guarantees — a page a protected policy says to
review is queued for your approval whichever agent proposed the edit.
- Codex — if you have a ChatGPT subscription, Wikai runs the Codex CLI
you already have, the way it runs Claude Code. No key to manage; sign in
once with
codex login.
- OpenRouter — one key, hundreds of models. Wikai lists the ones that
can call tools and leaves the rest out.
- llama.cpp and vLLM — a model you run yourself, on this machine or
one you own. Point Wikai at the server and the wiki answers with the
network off. Settings says, in one sentence, where your pages are going,
because that depends on the address you typed — and it warns you when a
server on the open internet has no key.
- Claude Code and the Anthropic API, as before.
Each agent has its own tab under Settings → Agents, with its brand mark,
what it needs from you, and a status line. The three least-exercised carry a
small beta tag. A Grok tab is there too, marked coming soon, so you can
see where it will go.
See the chat sidebar and
privacy & data for what leaves your machine with
each.
The model, beside the chat
- Pick the model where you chat. A dropdown next to the agent picker
lists what that agent offers — Codex reports its catalog, OpenRouter and
the local servers are asked for theirs, Claude Code takes an alias like
sonnet or opus. Leave it on Default and the agent picks, which is the
one choice guaranteed to be inside your plan. Wikai remembers your model
per agent.
- Reasoning effort for Codex, from the ladder each model actually
supports — a wrong value would cost you a failed turn, so you can only pick
a right one.
- “Last used” as your default agent. New chats open on whichever agent
you used last, across all your wikis.
- A reopened chat goes back to the agent it ran on, which it did not do
before — a Codex chat reopened with the picker on Claude Code sent its
next turn to the wrong agent.
Research is something you start
An agent that decides, on its own, to open a research project rewrites a
chat’s purpose and puts a file in your wiki. It can now only propose one:
you get a banner, and nothing is created until you accept. Asked a plain
question, an agent answers it.
The daily note understands time
- Today is a timeline. Captures sit at the minute you made them on a
vertical rail; hand-written prose sits between. Click an hour and it opens
in place with a composer in the gap, so a note at 14:20 lands at 14:20.
- Intent and Reflection — a section at the top of the day for what you
mean to do, and one at the bottom for what happened.
- A capture can become a section: promote a one-liner into a heading
with a body when it turns out to be a meeting, not a remark.
- Long entries fold, and a quiet stretch of the day folds into a wave
rather than a wall of empty hours.
- Summarise an entry. Ask the agent for a summary of a long entry; it is
offered beside the entry, grounded in your wiki, and written only if you
take it.
- Capture while the day is open in the editor — the capture lands where
it belongs without fighting your unsaved text.
Follow-ups on today’s plan
- Star a checklist onto today. A star in any checklist’s header puts
that list under today’s Intent — as a reference, not a copy. Tick a box
there and the source page is updated; the follow-up stays the follow-up.
- Plan a week by starring the days you mean to work on something, from
a small calendar in the same button.
- Done items are struck through, not hidden, from one definition of
done, so the plan and the source agree.
- Take a checklist off the plan from the day itself.
Elsewhere
- The connections panel sets its floor from your wiki, so it stops
showing the least-related pages as if they were related.
- Block embeds render as references in the editor, with a chip that
shows the label rather than the block id.
- Codex conversations continue past the first turn. The first build
could not resume a Codex thread; fixed before release, and the fix keeps
the sandbox on for resumed turns.
- Creating a research document no longer cuts off the agent mid-sentence,
and pressing Stop no longer reports an error.
- The Anthropic API backend can now use concept search.
Fixes: a failed Codex turn reports the actual reason instead of “exited with
status 1”; the chat header’s dropdowns keep their width when you switch
agents; and the settings page no longer allows Automatic daily-note
enrichment on any metered agent, while allowing it on a local model, which
is free.