Tagent logotagent
Tagent
v0.31.2 is out — navbar machine stats & a context bar for every provider

The coding agent that lives in your terminal, driven from your browser.

Tagent runs the engine on your machine — loop, subagents, MCP servers, plugins, skills, memory, GitHub sync, permissions. One self-contained binary (web GUI included) for Windows, macOS and Linux. The TUI is the interface (terminal or phone), the web GUI is the companion. Open source, BYOK.

# 1 — download one file for your platform (see Download)
# 2 — run it. that's the whole install
./tagent start ~/my-project        # full TUI
./tagent start --web-gui           # TUI + browser GUI
Tagent web GUI on desktop — chat, file tree, editor and terminal
Tagent mobile layout — files tab
Files & editor on a phone
Tagent mobile layout — chat tab
Chat with tool cards on a phone

Same engine, phone-sized.

Install Ubuntu with UserLAnd from the Play Store (no root), clone Tagent, and open localhost:4020 in Chrome. Bottom tabs: Chat, Files, Terminal, Memory, Skills.

Phone setup guide

Everything an agent needs

A complete loop with the safety rails on — inspired by the best of opencode, rebuilt as an open platform you fully control.

new

GitHub sync

Nothing requires an account — but log in (tagent auth) and tagent offers to sync the current project to a private GitHub repo. Continue on any machine with tagent clone.

new

Smart cache

The token economist — unchanged-file re-reads return a stub, read_files batches 12 paths in one turn, @path mentions attach files inline, and old tool outputs auto-compact. tagent cache shows the savings.

new

MCP servers

Plug the Model Context Protocol straight in — Context7, filesystem, memory, sequential-thinking or any stdio server. One command in the TUI, tools become native.

Agentic loop

Prompt → model → tool actions → results → repeat. Native function-calling, turn budgets, interrupt & steer any time.

Subagents

The task tool spawns isolated explore/general agents with their own budget — research while you keep building.

Permissions

Every write asks first — arrow-key prompt in the TUI, dialog in the GUI. Per-tool ask/allow/deny, remembered once, per session, or always.

Memory & skills

AGENTS.md + durable facts injected into every prompt; SKILL.md playbooks with progressive disclosure.

Multi-provider BYOK

Z.ai, OpenAI, Anthropic, Google, OpenRouter, Groq, Ollama — or any OpenAI-compatible endpoint. Your keys stay local.

new

Plugins

Hook into the loop, add custom agent tools and slash commands — hot-reloading .mjs files in .tagent/plugins/.

Web & browser tools

web_fetch and ddg_search work with zero API keys; an optional Playwright browser tool gives the agent eyes.

Runs on a phone

Ubuntu via UserLAnd (no root) — the GUI has a dedicated mobile layout with bottom tabs. localhost:4020 in Chrome.

TUI — the primary interface

interactive, app-like, arrow keys

  • Arrow-key menus everywhere — model picker, sessions, permissions, y/N confirms
  • Claude Code-style: transcript in your terminal's own scrollback, rounded editor box at the bottom
  • Scroll naturally — mouse wheel, touch on a phone, shift+pgup, tmux copy mode
  • Slash commands: /mcp, /plugins, /update, /model, /sessions, /auth… · @file mentions · ? shortcuts
  • Answers render as markdown — headings, code blocks, lists, tables
  • Hint row under the editor (model · tokens · ⎇ repo), ESC stops the run
  • Streams tokens live, queues your next message mid-run

Web GUI — the cockpit

same engine, same sessions, browser

  • Chat with tool cards, streaming, live todos and subagent timeline
  • File tree, editor, terminal — workspace in a tab
  • Settings: providers, MCP servers, plugins, permissions
  • GitHub login & project sync as dialogs, not terminal flags
  • Mobile layout with bottom tabs; live relay to share a session

One process. No cloud required.

The daemon is a single Bun process: agent engine, websocket API, and the pre-built web GUI served as static files. Nothing leaves your machine — API keys live in ~/.tagent/config.json, sessions in your workspace. Remote relay mode is on the roadmap, never required.

  • Bun runtime — no native compilation, no node-gyp
  • GUI ships pre-built — git clone & run, no Next.js toolchain
  • localhost-bound by default; LAN mode behind a flag
  • Self-updating — checks on startup, y/N to update in place
┌─ Browser (desktop or phone) ──────────┐
│  chat · tool cards · permissions      │
│  file tree · editor · terminal         │
└──────────────┬─────────────────────────┘
               │  websocket @ /socket
┌─ Tagent daemon (Bun, localhost:4020) ▼┐
│  agent loop · tools · subagents        │
│  sessions · checkpoints · permissions   │
│  mcp servers · plugins · GUI            │
└────────────────────────────────────────┘
      │                    │
  your keys            your files
  (BYOK, local)     (workspace-jailed)

Take the agent for a spin

The repo includes a demo workspace with planted bugs — a five-minute way to watch the loop read, patch, and verify real code.