Agents on your account

The agents on the AI agents page are the ones running in your terminals. This page is about the other kind: an agent that belongs to your account, runs on a schedule, and works when nobody is watching.

Open Dashboard → Agents and pick one to start talking to it. Everything else on this page lives in its settings: click the agent's name at the top of the conversation to get there.

Talk to itCopy Anchor Link

A conversation with an agent keeps its context, so a follow-up can say "now do the same for the other branch" without repeating itself. To keep one beside your terminals, open the command palette in a session and choose Open Agent Chat. It opens as a tile you can split and move like any other.

While a turn starts, the line under the conversation says what is happening before the model does anything: the agent host starting, connectors being mounted, then Working once it has your message. Type / in the box for the commands a conversation takes, among them:

An agent can ask you a question without stopping to wait for the answer. It keeps working, and the question waits in the conversation until you get to it: answer it, press Later to keep it and your draft for another time, or Dismiss it. /questions lists everything still open. An answer goes back into the conversation it came from, even after that run has finished.

Outputs: what a run left behindCopy Anchor Link

When a run finishes having written files, such as a long tool output it could not fit in the conversation or what a subagent wrote, a line under the conversation says so: The 3 files this run left. A run that left nothing adds no line. Press it to list the files, then press a file to read it in place. In a tile, Open as tile opens it beside the conversation, where a long log is easier to read.

Past conversationsCopy Anchor Link

Every conversation an agent has had stays reachable. Type /resume to choose one from a numbered list, newest first, or /resume 2 to go straight to the second newest. /continue opens the most recent one, and /new starts fresh. /rename Night shift names the conversation you are in. /rename on its own puts a name the agent suggests in the message box for you to keep or edit, and Enter names it. To keep the current name, press Esc, or on a phone, Backspace at the start of the box.

Above the conversation (on a phone, under Agent options), Conversations lists the same history with how many turns each had and when it started. New conversation sits at the top of it. A conversation marked cannot be continued can still be read. About this history says what the list holds: saved messages and summaries, so some subagents, jobs and images may be missing. An agent that has run more than 20 times shows only its last 20 turns.

SubagentsCopy Anchor Link

An agent can hand part of a task to subagents that work alongside it. Each one appears as a row under the reply that started it, with its own animal face, the task it was given, and its state: running (with how long it has taken), done, failed or stopped. A subagent that has a question for you says asks you. Press a row to read that subagent's own conversation.

While any of them is still working, their faces sit stacked in the tile's header, up to three with a count for the rest. Point at the stack to see each one's task. A subagent wears the same face in the thread and the header, so you can tell which is which. The stack goes away once they have all finished.

Its public pageCopy Anchor Link

Every agent has a page of its own at /u/<your username>/<slug>, where the slug is the short name the agent was created with. View public page, at the top of the agent's settings, opens it. Who can open it is the Who can see it setting under Behaviour: Only me, Friends, or Anyone signed in.

When it is shared, the page shows the agent's persona, model, how much it may do on its own, its skills, its schedule including what you told each job to do, and its recent runs with what each one wrote and which tools it used. Failures and refused tools stay with you. Visitors have to sign in first, and they can only read: there is no conversation on the page and nothing they can start or stop. To anyone it is not shared with, the page says the agent isn't available.

Try it before you schedule itCopy Anchor Link

On an agent's settings page, Run now runs it once. Leave the box blank to run whatever its schedule would have run, or type a one-off instruction. The run appears under History with what it did, so you can see how it behaves before you let it loose at 7am.

Skills: the things it knows about your workCopy Anchor Link

A skill is a short document that tells the agent how something is done here: how your nightly failures are usually read, what a release note should say. Add one by writing it in the box, or paste a link to a SKILL.md file and taiku will fetch it.

Skills are per agent, and you can switch one off without deleting it. A skill pack is a set of skills taiku ships and keeps reviewed; adding one is a single press, and each pack says what it changes about the agent before you take it.

A pack is copied onto the agent when you add it, so a later taiku can ship newer text than your agent has. The pack then says Out of date, and More shows what changed skill by skill. Update rewrites only the skills you have not edited, keeps each one on or off as you left it, and removes any skill the pack no longer ships.

When it may use a skillCopy Anchor Link

An agent reads a skill only when a task calls for it. Each skill on the agent's page has a choice beside it:

• Use freely: the agent reads it without asking.
• Ask first: the run stops and asks you the first time the agent reaches for it. Always allow remembers the yes, and the agent's page lists those under Skills loaded without asking, where you can take one back. A scheduled run, or any run nobody is watching, cannot use it.
• Never use: the skill is turned off and the agent is not told it exists.

Under Skill rules, a rule like pdf-* decides every skill it matches, including ones added later. The most specific rule wins, so a choice made on one skill beats pdf-*. Settings → Agent holds Skill rules for every agent, used where the agent has no rule of its own, and a skill no rule names is used freely. Each skill's choice shows what those rules decide.

Plugins: everything a marketplace ships, in one pressCopy Anchor Link

A plugin is a bundle someone else maintains, holding slash commands, connectors and more. Add one by its address and the repository its marketplace lives in:

code-review@claude-plugins-official
anthropics/claude-plugins-official

Nothing is installed at the moment you press Add. taiku records that this agent should have the plugin, and the machine that runs the agent installs it at the start of the next run. Until that happens the row says "Installs on the next run"; afterwards it shows the version that actually landed, which is how you can tell the difference between "I asked for this" and "this is really there".

What a plugin gives the agent:

• Its commands. The agent can use them the way it uses any built-in one.
• Its connectors. Any MCP servers the plugin ships come along with it.

What it deliberately does not give:

• Its skills. Skills reach an agent only when you add them yourself, above.
• Its extension modules. Code that would load before the agent has done anything is not run, on any agent.

Turning a plugin off takes it off the machine at the next run rather than leaving it installed and ignored, and removing it does the same and forgets where it came from. Everything a plugin does runs inside that agent's own workspace, on that agent alone. Plugins on one agent are invisible to your other agents and to everyone else.

Connectors: the tools it can reachCopy Anchor Link

A connector is an MCP server the agent may call while it runs. Add one by name, point it at an address, and store its credential once. taiku keeps the credential encrypted and never hands it back to the page; the row only tells you that one exists.

Connected accounts work the same way from the other direction: link an account once, then name which of them this agent may touch.

ScheduleCopy Anchor Link

Give the agent a cron expression or an interval and a timezone, and it runs without you. 0 7 * * * in America/Los_Angeles stays at 7am local through daylight saving, and it stays at 7am for you when you travel.

A schedule can name specific skills, which is how one agent can triage the build in the morning and write release notes in the afternoon without loading both sets of instructions into every run.

Behaviour: how much it may do on its ownCopy Anchor Link

Autonomy decides what happens when the agent wants to write a file or run a command. The strictest setting refuses anything it was not given, and the most permissive lets it act without asking. In between, taiku answers on your behalf using the list of things you allowed, and the run reports what it could not do rather than doing it.

Whatever you choose, an agent's run is confined to its own workspace.

Where it runs is either taiku's agent host, shared with the other agents there, or Its own box, a machine nothing else runs on. The box is made the first time the agent runs, and runs wait for it to come up, and wait again while it is asleep or being rebuilt, for up to about 12 minutes; the line under the setting says what state it is in and follows the box as it changes. An agent set to its own box never quietly falls back to the shared host. If a box cannot be made, has failed or is being deleted, the run is not started, and History says why and what to change. A run still waiting after about 12 minutes ends as timed out.

When a run on its own box ends, taiku keeps a copy of the conversation's history, so the next turn can continue it even if the box was deleted or rebuilt in between. The copy holds what the agent saw, including what its tools read and printed. It stays with your account, is never listed among the outputs, and is deleted when you delete the conversation, the agent or your account. It is kept in read-only and plan mode too: it is taiku's record of the conversation, not something the agent changes.

Autonomy decides what the agent asks you about before its tools act, and taiku checks each request the agent makes to it, from a tool or from a program it runs. In read-only mode the agent changes nothing: it can't write files, message another agent, publish or file a document, change a work item or act in a connected account. Plan mode is the same, except that the agent publishes its plan as a document you can comment on.

Summarize older turns at is how full a conversation may get before taiku shortens it. When a reply takes it past that share of the model's context window, the agent keeps the most recent part word for word, replaces the rest with a summary, and your next message carries on from there. The default is 80%. A lower line compacts sooner and more often; a higher one keeps more of the conversation as it was, for longer.

Whose login paysCopy Anchor Link

An agent on its own box can reach Claude through your Claude subscription once you have signed in to Claude under Settings → Agent. The login stays with taiku and is never copied onto the agent's box, and a box anyone else has joined stops getting it. A Claude agent on its own box runs Claude Code, which is what lets the login cover every Claude model, Sonnet and Opus included. If the box cannot run Claude Code, a run your login would pay for is not started, and History says why; with an Anthropic API key added, it runs on that key instead.

An API key is the alternative, per provider. Add one under Settings → Agent → Model providers and that provider's calls are billed to your account with it. Each provider's row says which key is in use.

With neither, the agent uses the key this taiku holds, within the model budget your plan includes. taiku counts that spend itself as each model call passes through, so the number is the same whatever the box reports.

Under Run now the agent's page says what a run is likely to cost and who pays: your subscription, your key, or your plan's budget, plus the box's hourly price while it is up. Once a run finishes, its cost is on its line in History.

MemoryCopy Anchor Link

Agents keep what you tell them across runs: what your build server is called, how you like pull requests, a decision you made last week. The agents you made yourself all read the same memory. An agent installed from a package reads and changes only the memory its package names, and it can keep notes about the repository it works in, which agents installed the same way read when they run there. Type /memory in an agent's chat, or open Settings → Agent, to see what is kept, turn memory off, or forget a file or all of it. Notes kept for a repository are listed under its name, and forgetting all of it includes them.

That includes runs of your agents on your own computer. taiku turns Claude Code's and Codex's own memory off for them and gives them this memory instead, so forgetting something here forgets it for those runs too. A taiku agent --local session in your terminal still keeps Claude Code's or Codex's own memory.

Where the settings areCopy Anchor Link

Everything above, apart from model keys, lives on the agent's own page, because it travels with the agent to every device you use. Settings → Agent holds your model keys and skill rules, which every agent of yours shares, and the things that belong to the device in front of you: whether terminals show agent badges, and how you are alerted when an agent needs an answer.