Assistant
The Assistant is Claude Code or Codex, opened in Daintree’s sidebar and wired into the app. Ask for work in a sentence. It looks up Daintree’s runbook for the job, creates the worktrees, starts the agents and tells you as each one replies.
Runs on the plan you already pay for. Claude Pro or Max through Claude Code, ChatGPT through Codex. No Daintree account and no Daintree model in the path.
you start an agent on each of the three auth issues, tell me when they’re done
▸ Runbook: Fan out disjoint work across parallel agents
▸ 3 worktrees, 3 agents launched with notices on
issue-412 · issue-415 · issue-418
◂ issue-415 replied: “Callback fixed, tests pass.”
412 and 418 still working
you have claude check 415 before I look
▸ Runbook: Verify an agent actually did the work
▸ Reviewer launched in issue-415
Two engines
Choose one the first time you open the Assistant. Both follow the same runbooks through the same tools. What changes is the model, the pace, and whose plan it bills.
Codex · your ChatGPT plan
Answers almost as you press Enter
Codex runs the Assistant on GPT-6 Luna, the smallest and quickest of OpenAI’s GPT-6 models. That suits orchestration, where most turns are a runbook search, a tool call and a short reply.
Claude Code · your Claude plan
Steady through long, branching jobs
Claude Code runs it on Sonnet, the middle of Claude’s range, on your Pro or Max plan. Quick enough for tool-heavy turns, strong enough to hold a competition, a debate or a multi-agent review together.
One request, end to end
The Assistant finds the runbook first, then drives Daintree through its local tools. Each agent gets its own worktree. Nobody polls: Daintree sends each reply to the Assistant the moment that agent says it is done.
You
Start agents on the three open auth issues and tell me when they’re done.
All three replied. #412 and #415 are ready to review; #418 is asking about copy.
Codex
GPT-6 Luna your ChatGPT plan
daintree-runbooks
“start an agent on each of these issues and tell me when they finish”
daintree 127.0.0.1
Core tool set: worktrees, agents, prompts, replies
Agents, one worktree each
All three replied. #412 and #415 are ready to review; #418 is asking about copy.
Same runbook, same tools, same agents. Only the Assistant’s own turns change pace with the engine, and the timing is illustrative. Daintree’s third server, daintree-docs, answers “how do I” questions and sits outside this route.
Runbooks
Before it acts on a task, the Assistant searches Daintree’s runbook catalog with one sentence saying what you asked for. When a procedure matches, it comes back with which tools to call, in what order, and when to stop and ask you. The Assistant follows it, along with the foundation runbooks it builds on.
~40 procedures in the catalog. A few of them:
What people ask it
The jobs you used to do by hand across a wall of agent terminals, and the runbook each one routes to.
you Have Claude and Codex both fix this, then judge each other.
Runbook Orchestrate a competition or peer‑judging
Launches the contestants side by side, collects their answers, asks each to judge the others and reports the tally with the reasoning behind it.
you Which agents are waiting on me?
Runbook Answer a waiting agent
Reads the exact question each blocked agent is asking. Tell it the answer and it picks the matching dialog option by its label.
you Give each of these worktrees its own database and port.
Runbook Give each worktree its own database, port and local environment
Helps you set up and verify per-worktree environment scripts, with your go-ahead, so the worktrees can run the app side by side without fighting over one port.
you Let me know when the build finishes.
Runbook Watch an outcome unattended
For a build running in an agent’s terminal, asks Daintree to notify it when that agent stops. If what you asked for has no real trigger behind it, it says so rather than pretending to watch.
You stay in control
Daintree’s own server decides which of its tools the session can call, and refuses the rest. Actions that ask for confirmation still stop for you, whichever set is on.
Worktrees and agents. Create a worktree, launch one agent or several, send prompts, wait for replies, read terminals, interrupt, rename, move and close them, and delete a worktree the session created.
Core, plus recipes, starting work on an issue, project checks, pull request, issue and CI reads, git activity, CopyTree context, resource lifecycle, restarting terminals, skills and diagnostics.
Neither set exposes git or forge writes, file edits or settings changes through Daintree. Change the set under Settings, Assistant; it applies to the next session. How tool sets work.
Cost and data
The Assistant is the CLI you installed, signed in the way you signed it in. Daintree adds tools, not a model endpoint.
Conversation
you → Claude Code or Codex → Anthropic or OpenAI
Billed to your Claude or ChatGPT plan, or your API key, exactly as in a terminal.
App control
assistant → daintree on 127.0.0.1
Worktrees, terminals and agents are reached through a local server that never leaves your machine.
Runbook search
one-sentence task summary → assistant.daintree.org
The Assistant is told to leave out names, numbers, file names and pasted output. Daintree attaches no token or ID. Turn off Follow runbooks and it stops.
Docs search
question → daintree.org/api/mcp
For “how do I” questions. Turn off Search documentation and it stops.
Before you rely on it
It can misread a terminal, pick the wrong worktree or stop early. A runbook narrows the path; it does not make the model right. The diff is still the evidence.
Claude Code blocks edits at its tool layer. Codex has no bundled deny list, so its session relies on its prompt, and the tool set is the boundary that is actually enforced.
Its tools live in the running app. Hide the panel and the Assistant hibernates after a while; reopen it and the same conversation resumes.
Claude Code and Codex. Copilot CLI is wired but held back as experimental. The Assistant can launch any of Daintree’s agents as workers; only these two can be the Assistant.
Open the Assistant, pick Claude Code or Codex, and hand it the first job you would otherwise have ferried between panels yourself.