TheOrchard.
Claude Code multi-agent supervision
without the wall of terminals.
Coordinate multiple Claude Code agents from one calm, local-first workspace. Keep parallel work visible, surface only actionable escalations, and preserve a durable audit trail.
The agents got autonomous.
The interfaces stayed loud.
The moment you run more than one autonomous agent, the tooling turns against you. Every dashboard lights up like a switchboard. Every tool call pings. Twelve terminal panes scroll faster than you can read. So you do the rational thing and stop looking — right up until the one agent that is actually blocked sits waiting, and nobody notices.
The problem was never the agents. It was an attention model that treats work as news. TheOrchard flips the rule: an agent quietly working is not an event. Only a blocked, broken, or escalating agent earns your attention — and when one does, it is impossible to miss.
Agents that have what they need are silent. The one that needs you is unmissable.
Five ideas,
one calm system.
A living ecosystem, not a task queue. A team channel, not a wall of terminals. One honest signal, not a switchboard.
Trees — one agent, rooted in a directory
A Tree is a single Claude Code session with a Plot: the directory it works in, its sandbox. Plant one, give it a Seed — a reusable role of system prompt, model, and autonomy — and it goes to work. Plant the same Seed again whenever you need that role.
Groves — teams that coordinate and escalate
A Grove is a team of Trees sharing a channel. They message each other, split work, and escalate to you through a Slack-like interface. You supervise a whole team from one conversation instead of babysitting a wall of terminals.
The calm attention model — work is not news
Every signal has exactly one honest meaning. A pulsing dot means an agent is working and needs nothing. An orange badge means blocked or broken — act. A quiet badge means unread. No tray at all means nothing needs you anywhere. Absence is the all-clear.
The Needs-You tray — triage a fleet from one place
Every blocked agent across all your sessions — permissions, errors, escalations — pinned to the top of the sidebar with inline Allow and Deny. When it is empty it disappears entirely, so its presence always means real work is waiting on you.
Local-first coordination — supervision stays on your machine
TheOrchard's coordinator, repository access, browser UI, and append-only SQLite event log run locally and bind to localhost. Claude Code still sends prompts and relevant context to the configured model provider under that provider's data terms.
Adjacent tools that don’t
solve this problem.
Lights up like a switchboard the moment agents start working. You learn to ignore it, then you miss the one alert that mattered.
Shows you raw output, not meaning. Twelve scrolling panes is not supervision — it is surveillance you cannot act on.
Runs jobs to completion on a trigger. It does not hold a live conversation, ask for permission mid-task, or let you steer.
Tells you what already happened. TheOrchard is where the work happens and where you intervene while it still matters.
A hosted supervisor stores orchestration state remotely. TheOrchard keeps coordination and its event log local, while each Claude Code session continues to use its configured model provider.
Plant, assemble, supervise.
Ship under your eye.
Plant a Tree
Root an agent in a directory, pick a Seed, and set its autonomy — Ask, Auto-edits, Plan, or Full autonomy. One honest amber warning on bypass, no nag dialogs.
Assemble a Grove
Group Trees into a team that shares a channel. Set autonomy per member, mixed-trust if you want it. They message each other, coordinate, and escalate to you.
Supervise from the tray
Let them run. A busy team stays silent until one agent genuinely needs you — then it surfaces in the Needs-You tray with Allow and Deny inline. Triage without opening a single session.
Ship under your eye
Every message, tool call, permission, and cost datum is an ordered event in the log. Bypass skips approval, never observability. Reload-proof: nothing re-alarms, nothing gets lost.
When TheOrchard is the
answer you’ve been looking for.
Three, ten, twenty sessions in parallel. Each one is fine alone; together they are a wall of noise you cannot triage.
Every tool call pinged you. You muted it. Now you miss the permission prompt that actually blocks real work.
Someone will ask what an agent did and when. You need an append-only record of every action, not a scrollback buffer.
TheOrchard keeps orchestration state and its event log local. Model traffic still follows the data handling of the Claude Code provider you configure.
What you get
out of the box.
No setup theater. Install, build, start — it binds to localhost, opens your browser, and the whole ecosystem lives on your machine.
Absence is the all-clear. When nothing needs you, the interface gets out of your way.
Bypass skips approval, never observability — every action a fully autonomous agent takes still lands in the append-only event log. A pending permission survives a reload into the tray; replayed history never re-badges. Nothing re-alarms, nothing gets lost.
Most tools compete on how much they can show you. TheOrchard competes on how little it has to.
Three reasons this
feels different.
Built by operators who run fleets.
TheOrchard grew from the team’s own need to supervise parallel AI coding agents without turning every tool call into an alert. The attention model is shaped by that operating constraint.
One idea, ruthlessly applied.
The event log is the spine. Every signal is a pure fold over an ordered log, which is why a pending permission survives a reload, replay never re-alarms, and the whole attention model is testable in plain Node.
Calm is the feature.
Anyone can add more panels and more badges. The hard, valuable thing is a tool that stays silent when it should and is unmissable when it must be. That restraint is the product.