termagent

A terminal AI coding agent — like opencode, codex, and claude code — rewritten in TypeScript on Ink with the UI component system of termcn. Runs on any POSIX shell including Termux (Android).

It speaks the OpenAI-compatible Chat Completions API (streaming, tool calling, reasoning deltas, usage reporting), so it works with OpenCode Zen (default), OpenAI, OpenRouter, Together, Groq, any vLLM/llama.cpp/Ollama/LM Studio endpoint.

startup	chat + tool call + usage

tool approval (risk badge)	model picker

live theme hot-swap (tokyo-night)	help

Features

• Full-screen Ink TUI on the termcn design system: ❯_ header with status badges, scrollable chat feed, bordered composer, hints footer
• Real termcn components: ToolCall (live braille spinner + duration), ToolApproval (risk badge, [y] Approve / [n] Deny / [a] Always allow), ModelSelector (grouped, arrow-key), TokenUsage (⟨ 8.4k in / 512 out · model ⟩), ChatMessage, ThinkingBlock, DiffView, Badge, Panel
• 17 termcn themes, hot-swappable with /theme (persisted), --theme, --list-themes
• Tools: bash, read, write, edit (renders a colored diff), ls, glob, grep (ripgrep if present), todo
• Approval gates with risk classification: rm -rf, curl | sh and friends escalate to HIGH
• Token usage per turn via stream_options.include_usage; providers that reject the field are retried without it automatically
• Reasoning support: reasoning_content streams collapse into a ✻ Thought block (TERMAGENT_SHOW_REASONING=1 prints the full text)
• Session save/resume (/save, /load, /sessions) — same JSONL format as the earlier Python build, so old sessions load
• Stale-model auto-heal: if a saved Zen model id disappears upstream, the app switches to an available one and says so
• One-shot mode: termagent "explain this repo" (plain streaming, no TUI)
• Zero runtime dependencies beyond ink + react (+ cli-spinners, string-width)
• ASCII fallback (TERMAGENT_ASCII=1) for non-UTF terminals

Provider status: OpenCode Zen free tier

Zen's free models used to work with no key from any client. As of late 2026 the free tier is gated server-side to the official OpenCode client — third-party clients get 403 FreeTierError. termagent detects this and explains your options instead of dumping JSON:

• Set a Zen API key (opencode.ai/auth), then /models lists all current Zen models.
• Or point at any other OpenAI-compatible endpoint (see below).

termagent does not spoof OpenCode's client headers to bypass the gate.

Install

Requires Node 20+ (pkg install nodejs git on Termux).

git clone https://github.com/loak7993-code/termagent && cd termagent
npm install
npm run build
npm link                # or: npm install -g .
termagent

One-liner installer (Termux + Linux):

curl -fsSL https://raw.githubusercontent.com/loak7993-code/termagent/main/install.sh | bash

Configure

Pick one:

# 1) Environment variables
export TERMAGENT_API_KEY="sk-..."        # or OPENAI_API_KEY
export TERMAGENT_BASE_URL="https://api.openai.com/v1"
export TERMAGENT_MODEL="gpt-4o-mini"

# 2) Interactive config (saved to ~/.config/termagent/config.json)
termagent --config

Other providers:

termagent -b https://opencode.ai/zen/v1 -k $ZEN_KEY -m glm-5.2
termagent -b https://openrouter.ai/api/v1 -k $OPENROUTER_KEY -m "anthropic/claude-3.5-sonnet"
termagent -b http://localhost:11434/v1 -k dummy -m llama3.1    # local Ollama
termagent -b http://localhost:1234/v1 -k dummy -m my-model     # LM Studio / llama.cpp

Usage

termagent                      # interactive TUI
termagent "list the python files"   # one-shot
termagent -c /path/to/project  # working directory
termagent -a                   # auto-approve tool calls
termagent -s 20260721-214103-cc34ae  # resume a session
termagent --theme tokyo-night  # pick a UI theme
termagent --list-themes        # all 17 themes

Inside the TUI: type a task, or /help for all commands (/models, /model, /theme, /approve, /save, /load, /sessions, /cwd, /tools, /sys, /clear, /paste, /config, /exit). Arrow keys scroll the feed; overlays close with esc.

Configuration reference

Variable / flag	Purpose	Default
TERMAGENT_API_KEY / -k	API key	empty
TERMAGENT_BASE_URL / -b	OpenAI-compatible base URL	https://opencode.ai/zen/v1
TERMAGENT_MODEL / -m	Model id	big-pickle
TERMAGENT_THEME / --theme	UI theme	opencode
TERMAGENT_SHOW_REASONING	print full reasoning text	off
TERMAGENT_ASCII	ASCII-only symbols and borders	off
TERMAGENT_CONFIG_DIR	config + sessions directory	~/.config/termagent
--max-turns	max agent turns per user message	24
-a / --auto-approve	approve tool calls without asking	off

Build, test, develop

npm install
npm run typecheck     # tsc --noEmit, strict
npm test              # vitest: 29 tests (agent loop vs mock SSE server, tools, themes, UI)
npm run build         # tsc -> dist/
npm link              # installs the `termagent` binary

To regenerate the UI screenshots: run the app inside tmux and capture panes (see docs/previews/ for the verified outputs).

Architecture

src/
  cli.tsx            arg parsing, TUI bootstrap, one-shot mode
  app/
    App.tsx          app-shell layout, feed rendering, overlays
    store.ts         observable store bridging the agent loop to React
    commands.ts      slash commands, model/theme pickers
  agent/
    agent.ts         tool-calling loop (timing, usage, reasoning callbacks)
    client.ts        fetch-based SSE client (chunked tool-call accumulation, usage)
    config.ts        JSON config + env overrides
    zen.ts           OpenCode Zen catalog + live /models fetch
    session.ts       JSONL session save/resume (compatible with the Python build)
    tools/           bash, read, ls, glob, grep, write, edit, todo + registry
  components/ui/     vendored termcn components (shadcn registry, MIT)
  hooks/ providers/  vendored termcn hooks and theme/unicode providers
  lib/
    terminal-themes/ 17 termcn theme palettes (verbatim from the termcn registry)
    themes.ts        theme registry + resolution
tests/               vitest suite incl. a scripted OpenAI SSE mock server
docs/previews/       verified screenshots

Known limitations

• Zen free tier is OpenCode-client-gated upstream (see above); the catalog covers chat-completions families only — Claude/Gemini/GPT-5.x on Zen use /messages or /responses and are excluded.
• The model picker shows the first 12 models; /model <id> sets any model directly.
• Reasoning content is summarized as a duration note unless TERMAGENT_SHOW_REASONING=1.
• One-shot mode is plain streaming output (no TUI), like claude -p.

History

v0.2.0-python (tag) is the previous pure-Python implementation (rich-based, termcn-styled). The TypeScript build replaces it; session files are compatible.

Credits

UI components and themes from termcn (MIT), rendered with Ink. termagent is MIT.