tars

README

TARS

TARS is a local-first AI project autopilot.

Unlike Claude Code, Aider, or Cursor, which mostly operate at the file and conversation level, TARS manages long-running projects autonomously: it plans phases with you, executes backlog work inside each phase, coordinates tools and worker agents, and only brings you back for approvals or real blockers. 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. Plan — Collects requirements through a short interview and turns them into a phase plan
2. Phase Loop — Builds a backlog, selects the next task, executes it, evaluates the result, and replans when needed
3. Capabilities — Combines built-in tools, skills, MCP servers, web research, and worker agents in one runtime
4. Human-in-the-Loop — Escalates at phase approvals and real blockers instead of asking for every routine retry
5. Dashboard — Live phase status, run status, pending decisions, blockers, and worker reports in a browser

tars init && tars serve
tars

Open the web console, then start a chat such as todo 앱 만드는 프로젝트 시작해줘.

Agent Runtime

• Browser-based operator console + 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
• Unified memory console: manage MEMORY.md, memory/experiences.jsonl, daily durable memory files, semantic memory artifacts, and the knowledge base from one page
• Obsidian-style knowledge base: durable markdown wiki notes, graph/index metadata, built-in KB CRUD tools, and explicit opt-in lookup from memory search
• 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
• MCP transports for local stdio servers and remote HTTP/WebSocket endpoints, with bearer or OAuth auth for remote servers
• 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 skills and MCP servers with manifest metadata, runtime gating, and default project profiles
• 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 web console
tars

tars tui is now a hidden, deprecated escape hatch for legacy debugging only. The supported paths are the web console and one-shot CLI commands.

Kick off a project from chat in the console, or use the CLI commands directly:

tars project activity <project-id> 20
tars project autopilot start <project-id>
tars project autopilot advance <project-id>
tars project autopilot status <project-id>

The recommended path is planning first, then controlled phase advancement. advance runs one synchronous autopilot step so you can inspect approvals, blockers, and replans explicitly, and status shows the current phase, run status, and next action without switching to the web console.

Project-linked cron jobs also inherit the project's tool allowlist now, so background workflows can use approved shell/file tools for examples like the bundled ops-service triage loop.

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 console: http://127.0.0.1:43180/console

The console now includes a dedicated Memory page at /console/memory for editing durable memory files, testing memory_search, and browsing or editing compiled knowledge-base notes. Legacy /console/knowledge links still open the same page.

Install trusted MCP packages from the hub:

tars mcp search
tars mcp install safe-time

Local stdio MCP servers 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

When you run TARS directly from a source checkout, build the embedded web console once before opening /console:

make console-install
make console-build

For live frontend work, run npm run dev inside frontend/console and start the Go server with TARS_CONSOLE_DEV_URL=http://127.0.0.1:5173.

Documentation

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

Status

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