tars

README

TARS

TARS is a local-first AI project autopilot.

Unlike Claude Code, Aider, or Cursor — which operate at the file and conversation level — TARS manages entire projects autonomously: it seeds a Kanban board from a natural-language brief, dispatches tasks to AI worker agents, reviews their output, retries failures, and keeps going until the project is done. All in a single Go binary running on your machine.

Key Features

Project Autopilot

The killer feature. Describe what you want to build, and TARS handles the rest:

1. Brief — Collects requirements through a short interview
2. Board — Seeds a Kanban board with todo / in_progress / review / done stages
3. Dispatch — Assigns tasks to AI worker agents (Claude Code CLI, Codex, or any gateway agent)
4. Review — Validates test/build results and GitHub Flow metadata before promotion
5. Retry — Auto-recovers stalled work without asking you to intervene
6. Dashboard — Live project status, worker reports, and PM notes in a browser

tars init && tars serve
# In the TUI:
> todo 앱 만드는 프로젝트 시작해줘

Agent Runtime

• Terminal client with a Bubble Tea TUI + local HTTP API (tars serve)
• Session lifecycle, transcript storage, and structured context compaction
• Agent loop with built-in file, process, scheduling, memory, and ops tools
• Built-in file tools with 2,000-line read pagination, continuation hints, and safe atomic writes
• Structured session compaction with identifier-preserving summaries, a safer recent-tail preserve policy, and manual /compact [instructions]
• Parallel read-only chat subagents through the built-in explorer gateway agent
• Semantic memory recall with Gemini embeddings (optional)
• Playwright-based browser automation

Extensibility

• Skill Hub — tars skill search, tars plugin install, and tars mcp install from a vetted registry
• Plugins — Bundle MCP servers, tools, and skills into installable packages
• Managed MCP Hub — Install checksum-verified MCP packages hosted in tars-skills
• Skills — LLM instruction files (SKILL.md) with companion scripts and runtime gating by plugin, binary, env, and platform requirements

Install

Homebrew:

brew tap devlikebear/tap
brew install devlikebear/tap/tars

Curl:

curl -fsSL https://raw.githubusercontent.com/devlikebear/tars/main/install.sh | sh

Quick Start

# 1. Initialize workspace and config
tars init

# 2. Set your LLM provider
export OPENAI_API_KEY="your-api-key"
# Or use Claude Code CLI: set llm_provider: claude-code-cli in config

# 3. Validate setup
tars doctor --fix

# 4. Start the server
tars serve --config ./workspace/config/tars.config.yaml
# Or as a macOS background service:
tars service install && tars service start

# 5. Launch the TUI client
tars

Kick off a project from chat, or use the TUI commands directly:

/project board <project-id>
/project dispatch <project-id> todo
/project autopilot start <project-id>

For read-heavy codebase research in chat, TARS can now fan out parallel explorer subagents and merge back compact summaries. The runtime defaults are:

gateway_subagents_max_threads: 4
gateway_subagents_max_depth: 1

Open the dashboard: http://127.0.0.1:43180/dashboards

Install trusted MCP packages from the hub:

tars mcp search
tars mcp install safe-time

Hub-managed MCP packages still respect mcp_command_allowlist_json. For example, a Node-based MCP package requires a config allowlist such as:

mcp_command_allowlist_json: ["node"]

Requirements

• Go 1.25.6+ (for building from source)
• LLM provider credentials, or a local Claude Code CLI install
• Optional: Gemini API key for semantic memory embeddings
• Optional: Node.js for Playwright browser automation

Build

make build-bins
bin/tars version

Documentation

• Getting Started
• Project Workflow Example
• Plugin and MCP Packaging Guide
• Contributing
• Changelog

Status

Pre-1.0.0 — Module path: github.com/devlikebear/tars