AgentTakt

AgentTakt is an MCP (Model Context Protocol) server and TUI tool. When an AI agent (an "Executor" such as Claude Code) sends a task execution plan over MCP, AgentTakt renders it as a node graph in your terminal. You review it with mouse and keyboard — move, add, and delete nodes, draw dependency edges, edit parameters — then approve, and the edited plan JSON is returned to the Executor for execution.

Claude Code (Executor)

│ stdio (MCP) your other terminal ▼ │ [agenttakt serve] ── Unix domain socket ──▶ [agenttakt (TUI)] MCP server review / edit / approve

Features

• Terminal-native — no web UI; everything runs inside your terminal
• Visual node editor — rounded nodes, dependency edges, and per-type coloring, powered by Textual
• Mouse-first editing — drag nodes to move them, draw edges between ports (rubber band), click to select and delete
• Safe approval loop — cycle detection (DAG guarantee) and other validations at the entry point, returning errors the agent can self-correct

Requirements

• Python 3.10+ (recommended: uv)
• A terminal emulator with mouse reporting (iTerm2, WezTerm, kitty, Ghostty, ...)

Installation

If you have [uv](https://docs.astral.sh/uv/), no installation is needed. uvx agenttakt fetches and runs AgentTakt on demand, and the .mcp.json example below starts the MCP server the same way.

If you don't have uv, install AgentTakt once:

brew install ryoohshima/tap/agenttakt    # Homebrew

pipx install agenttakt # pipx

Installing is also handy for everyday use even with uv — you start the TUI by hand, so plain agenttakt beats typing uvx agenttakt each time:

uv tool install agenttakt

Agent Plugins

This repository is an Agent Plugins 1.0.0 package. In a compatible client, load the repository root as the plugin directory:

AgentTakt/

├── plugin.json # Portable plugin manifest ├── mcp.json # MCP server configuration └── LICENSE

The plugin requires uvx on PATH and runs the published agenttakt package from PyPI. Its uv cache is stored under the client-managed PLUGIN_DATA directory. Start the TUI separately with uvx agenttakt before using the tools. For request_approval, configure a sufficiently long tool timeout in your client (for example, 30 minutes); the portable MCP format has no timeout field.

The development .mcp.json is a separate client-native configuration that runs the local checkout. server.json supplies metadata for the MCP Registry.

Quick Start

AgentTakt runs as two processes: the MCP server, which Claude Code starts for you, and the TUI, which you start yourself in a separate terminal. The TUI is what displays the plan, so start it before asking the Executor for approval.

┌─ Terminal A: you ───────────────────┐   ┌─ Terminal B: Claude Code ───────────┐

│ $ uvx agenttakt │ │ $ claude │ │ │ │ │ │ ╭─ grep ───╮ │ │ > Plan the refactor, then ask │ │ │ pattern │───╮ │ │ me to approve it │ │ ╰──────────╯ │ │ │ │ │ ╭────▼─────╮ │ │ calls request_approval(plan) │ │ │ edit │ │ │ waiting for approval... │ │ ╰──────────╯ │ │ (blocked until you decide) │ │ │ │ │ │ [a] Approve [r] Reject │ │ │ └─────────────────────────────────────┘ └─────────────────────────────────────┘ ▲ │ ╰──────────────── Unix domain socket ────────────────╯

Running the TUI in the same session as Claude Code does not work. A stdio MCP server has its standard input and output reserved for protocol traffic, so the same process cannot also drive a full-screen terminal UI. That is why the two halves are separate processes talking over a Unix domain socket.

1. Start the TUI (in its own terminal)

uvx agenttakt           # if installed: agenttakt (short alias: agt)

An idle screen appears, waiting for plans from the Executor. Leave this terminal open. If no TUI is running when the Executor calls request_approval, the call fails with:

AgentTakt editor is not running. Ask the user to run "agenttakt" in a separate terminal, then call request_approval again.

On startup the TUI checks PyPI in the background and shows a notification when a newer version is available. Set AGENTTAKT_NO_UPDATE_CHECK=1 to disable the check.

2. Register the MCP server with the Executor (Claude Code)

Add the following to your project's .mcp.json:

{

"mcpServers": { "agenttakt": { "command": "uvx", "args": ["agenttakt", "serve"], "timeout": 1800000 } } }

[!IMPORTANT]

Setting `timeout` (milliseconds) explicitly is required. The request_approval tool blocks until the human finishes reviewing. MCP progress notifications do not extend client-side timeouts, so the default would cut the request off before approval. The example above sets 30 minutes (1800000). This does not apply to show_plan, which returns as soon as the TUI receives the plan.

3. Request approval from the Executor

When the Executor calls the MCP tool request_approval(plan, summary), the plan appears in the TUI as a node graph. Once the human edits and approves (or rejects) it, the result is returned as:

{ "status": "approved", "plan": { "...edited plan..." }, "reason": null }

See docs/schema.md for the plan JSON format and what to write in each node.

Display-only plans (show_plan)

show_plan(plan, summary) shows a plan in the TUI without waiting for approval — it returns {"status": "displayed"} as soon as the editor receives it. Use it when you just want visibility into what the agent is planning, in any mode (not only plan mode). The plan opens with a [view-only] header; closing it sends nothing back to the Executor.

Agents call request_approval naturally when the host is in plan mode, but they will not volunteer plans outside it. To encourage that, add an instruction like this to your project's CLAUDE.md (or equivalent agent instructions):

## AgentTakt

Whenever you formulate a multi-step plan — in any mode, not just plan mode — submit it with the AgentTakt show_plan tool so the human can see it as a node graph. Use request_approval instead when you need the human's approval before executing.

Note: a [view-only] plan occupies the editor until dismissed; a later request_approval waits in the queue behind it.

Debug mode (try it without MCP)

uvx agenttakt open examples/sample_plan.json --out edited.json

Loads a plan from a file, opens the editor, and writes the approval result to --out.

Key Bindings

Mouse: drag a node to move it; drag from a node's output port (●, right edge) and release on another node to create an edge.

Edges are drawn as braille Bezier-like curves by default. If they render poorly in your environment, switch to rounded orthogonal lines with --edges orthogonal.

Documentation

• Plan JSON schema — data model, node fields, what to write in type / data, and validation rules
• Changelog — release notes for each version

License

MIT