Conductor - visual workflow planner in Claude Code

A terminal UI for planning, running and understanding complex agent workflows inside Claude Code.

Don't hope that Claude one-shots your multi-step logic from a loosely worded description.

Try design → review → run instead.

/plugin install workflow-conductor --marketplace bruw/workflow-conductor
/conductor
# make sure you're in a large terminal window

Conductor has Claude draw a visual plan that thinks through the workflow step by step. You can read it in ten seconds and check the logic before Claude writes the script.

A plan costs about as much as Plan mode, and it can save a lot of tokens compared with a workflow done badly.

What

• Design: Draw it yourself, start from a template, or let Claude draft it from a description and refine it in plain words.
• Check: Common design mistakes are flagged on the graph before you run. Flags are warnings, not blocks: you can run anyway. It's your money.
• Run: Claude implements the spec as a Workflow script and runs it.
• Monitor: Progress, failures and token cost show up on the plan you made.
• Learn: Block kinds are named multi-agent patterns, and you can compare each design with the script it produced. That makes Conductor a practical way to learn how good workflows are built.

How

Type /conductor review this PR thoroughly. Claude drafts this and puts it on the board:

╭───────────────────────────────────────────────────────────────────── n6 ─╮
│ START                                                                    │
│ Review the current diff along several dimensions and keep only findings… │
╰──────────────────────────────────────────────────────────────────────────╯
                                     ↓  [+]
╭───────────────────────────────────────────────────────────────────── n1 ─╮
│ AGENT scout / scout diff                                             low │
│ Cheap and mechanical. Its list of dimensions is the work-list for the p… │
╰──────────────────────────────────────────────────────────────────────────╯
                                     ↓  [+]
┌╴ PIPELINE / review · once per: ×4 dimensions ╴╴╴╴╴╴╴╴╴╴╴╴╴╴╴╴╴╴╴╴╴╴╴ n2 ╴┐
╎ ╭───────────────────────────────────────────────────────────────── n3 ─╮ ╎
╎ │ AGENT / review                                                       │ ╎
╎ │ One reviewer per dimension.                                          │ ╎
╎ ╰──────────────────────────────────────────────────────────────────────╯ ╎
╎                                    ↓  [+]                                ╎
╎ ╭───────────────────────────────────────────────────────────────── n4 ─╮ ╎
╎ │ AGENT verify / verify findings                                       │ ╎
╎ │ 3 votes: each skeptic tries to refute a finding and defaults to ref… │ ╎
╎ ╰──────────────────────────────────────────────────────────────────────╯ ╎
└╴╴╴╴╴╴╴╴╴╴╴╴╴╴╴╴╴╴╴╴╴╴╴╴╴╴╴╴╴╴╴╴╴╴ [+] ╴╴╴╴╴╴╴╴╴╴╴╴╴╴╴╴╴╴╴╴╴╴╴╴╴╴╴╴╴╴╴╴╴╴╴┘
                                 all ⇓  [+]
╭───────────────────────────────────────────────────────────────────── n5 ─╮
│ AGENT merge                                                              │
│ Runs once over all surviving findings: merge duplicates, order by sever… │
╰──────────────────────────────────────────────────────────────────────────╯

Edit it directly in the UI, or ask Claude to edit it for you.

Then press r to script and run. Claude asks only about what is unresolved, states the plan with its worst-case cost, writes a real Workflow script from the project, and runs it. The board becomes the run:

╭───────────────────────────────────────────────────────────────────── n1 ─╮
│ AGENT scout / scout diff                                             low │
│ Cheap and mechanical. Its list of dimensions is the work-list for the p… │
╰─ ✗ 3/4 done · 1 failed · 20.4k ──────────────────────────────────────────╯
                                     ↓
┌╴ PIPELINE / review · once per: ×4 dimensions ╴╴╴╴╴╴╴╴╴╴╴╴╴╴╴╴╴╴╴╴╴╴╴ n2 ╴┐
╎ ╭───────────────────────────────────────────────────────────────── n3 ─╮ ╎
╎ │ AGENT / review                                                       │ ╎
╎ │ One reviewer per dimension.                                          │ ╎
╎ ╰─ ✓ 4/4 done · 30.6k ─────────────────────────────────────────────────╯ ╎
╎                                    ↓                                     ╎
╎ ╭───────────────────────────────────────────────────────────────── n4 ─╮ ╎
╎ │ AGENT verify / verify findings                                       │ ╎
╎ │ 3 votes: each skeptic tries to refute a finding and defaults to ref… │ ╎
╎ ╰─ ✓ 4/4 done · 40.8k ─────────────────────────────────────────────────╯ ╎
└╴ ✓ 8/8 done · 71.4k ╴╴╴╴╴╴╴╴╴╴╴╴╴╴╴╴╴╴╴╴╴╴╴╴╴╴╴╴╴╴╴╴╴╴╴╴╴╴╴╴╴╴╴╴╴╴╴╴╴╴╴╴╴┘
                                 all ⇓
╭───────────────────────────────────────────────────────────────────── n5 ─╮
│ AGENT merge                                                              │
│ Runs once over all surviving findings: merge duplicates, order by sever… │
╰─ · not run ──────────────────────────────────────────────────────────────╯

Cost

Workflow script estimates, in tokens.

Flow	Design	Hand-off (r)
Draw it yourself, press r	~0k: the board makes no Claude calls	~10–20k, what Claude spends writing any Workflow script, but from a full spec
/conductor <description…>, press r	~5–15k for the draft, like Plan mode	~10–20k
Draft, q for a review, fix, press r	~10–25k, like Plan mode plus a round of refinement	~10–20k

The parts

There are eight blocks:

• Work: agent, code, subworkflow
• Containers: pipeline, parallel, loop, branch
• Goal: start

Each block comes in several kinds. A kind is a common pattern that prefills the block's logic field:

Block	Kinds
agent	task · scout · check · verify · judge · critic · merge · classify
loop	custom · count · pass · dry · converged · target · budget
parallel	arms · each · sweep · attempts
branch	if-else · guard · route

Four templates to start from (Review, Research, Migrate, Judge panel), or a blank board.

Getting around

Conductor supports full mouse and keyboard navigation.

/conductor	the board: the live run, else the last project, else the start screen
/conductor <description…>	Claude drafts a new project from your words
/conductor min · max	fold the designer away, bring it back

Key	Does
w a s d	move the selection; with Shift, move the block
Enter	edit the block's fields
n x t	add a block, delete it, change its type or kind
1 2	step the model, the effort
o b	open a container as the view, back up a level
u y	undo, redo: fifty steps a project
r	implement and run
q	ask Claude to review the design
p	projects and templates
f	write the board over a project file that is no longer valid
?	every other key, and the mouse

Requirements

• Claude Code 2.1.296 or later, with the Workflow tool.
• A terminal 110 columns wide, in the fullscreen layout. Narrower than that, you get the compact strip and no designer; widen it and the designer opens by itself.
• Fullscreen for the mouse. The keys work without it.

Develop

It's MIT-licensed, so feel free to change it to fit your needs.

claude --plugin-dir /path/to/conductor

In the desktop app, where no flag can be given, the env of your user settings takes it:

{ "env": { "CLAUDE_CODE_PLUGIN_DIRS": "/path/to/conductor" } }

• bun install, then bun run check: the plugin validation, tsc and the tests.
• bun scripts/render-graph.ts --list names what it can draw, and draws it with no session.
• scripts/dev-session.sh starts Claude Code with the mod in tmux; scripts/send.sh sends it keys, scripts/capture.sh prints the screen and the debug lines.
• src/ is pure logic; hooks/ is what Claude Code loads.

What it touches

Conductor makes no network requests and reads no settings or credentials.

Area	What Conductor does	What it never does
Files	Writes only under the session's root: .claude/workflow-conductor/<name>.json, and, when you answer Save to repo after a project's run, a copy of the script that run ran to .claude/workflows/<name>.js, which the Workflow tool runs by name; <name> is always kebab-case. On a hand-off Claude passes the Workflow script inline; the Workflow tool keeps its own copy in the session folder, and Conductor remembers that path for the next hand-off.	Write any other file another tool runs, settings, CLAUDE.md, build or start-up files. Delete or rename a file: to remove or rename a project, remove or rename its files.
Programs	Nothing.	Run any program.
Prompts	Submits a prompt to your own session in two cases. r: Implement and run the "<name>" workflow project (Conductor hands it off: call mcp__workflow-conductor__before_workflow first). /conductor <description>: Draft a workflow project: <description>, your own words. q and the draft key only put a sentence in the prompt box for you to send.	Put file content or conversation text in a prompt it submits: only the project name, or your own words. Read the prompt box beyond whether it is empty.
Tools	Stands in for two tools, both its own, on every call: its tool.call hook returns the result, and no other tool runs in their place. It is the only way a mod serves a tool it adds. mcp__workflow-conductor__before_workflow (hooks/offer.tsx): Claude calls it before writing or running a Workflow script; Conductor asks you whether to design the workflow on the board first and answers with your choice; subagents and -p runs get the standard answer without a question. mcp__workflow-conductor__project (hooks/project.tsx): Claude calls it to read and edit a project (read, apply, write, attach); the project's goal, blocks and file paths reach Claude this way and in notes on the prompts you send while the board is open.	Answer another tool's call in that tool's place.
Claude's tools	Refreshes the board after a project file is written and tracks a started run. Refuses a write or edit of a project file while that project is running. Asks once whether to draw a Workflow run no project claims: a yes, or what you type instead, is added to the tool's result as a note to Claude.	Change Claude's own Write, Edit and Workflow calls otherwise.
Hooks	/conductor (command.run, on that command name alone): opens the board or the run view, folds or expands the designer, or starts a draft, the same whoever runs the command, you or other code. Subagent starts, model requests and turn ends (agent.spawn, turn.step, turn.complete): reads a workflow agent's id and label, a subagent's token usage and how its turn ended, to draw the run, and passes each on unchanged; the end of Claude's own turn after a run finished is when the save question is asked. The prompt box (prompt.edit): while the board is taking the keyboard back, one letter typed into the empty box goes to the board as a key instead. Prompts you send (prompt.submit): reads each for the hand-off sentence, and adds a note of board changes Claude has not seen as context. Run ends (tool.call on TaskStop, the task-notification row's UserMessage render): reads which run was stopped or ended. session.compact and session.end (on clear or resume): reset what Claude was told.	Hook any other command. Read a prompt, instruction or tool description in the run hooks.