acomplete

Inline Tab autocomplete for the prompts you type into terminal coding agents — Claude Code, Codex, or any CLI. As you type, a grey ghost-text suggestion appears right after your cursor. Press Tab to accept, keep typing to ignore.

❯ add a dark mode toggle␣that persists the preference in localStorage
  └ you typed ──────────┘ └─ grey ghost text, press Tab ──────────────┘

Why

Terminal agents own their input box — there's no plugin hook for autocomplete. acomplete wraps the agent in a pseudo-terminal, watches what you type, and renders context-aware suggestions inline. It works with any terminal program; it's tuned for coding agents.

Features

• Inline ghost text — dim suggestion at the cursor, not a popup.
• Context-aware — uses your partial prompt, the recent conversation on screen, and your installed Claude Code skills.
• One key — Tab accepts the whole suggestion, → accepts one word, Shift+Backspace undoes the last accept, typing dismisses.
• Rephrase in place — Ctrl+R rewrites your whole prompt to be clearer; Shift+Backspace puts the original back.
• Zero-setup auth — reuses your existing Claude Code login by default. No API key needed.
• Bring your own model — or plug in Anthropic / DeepSeek / OpenAI / Gemini.
• Never in your way — on any error it degrades to a plain transparent passthrough.

Install

npm install -g acomplete

Or straight from GitHub:

npm install -g github:dilyorm/autocomplete

Requires Node ≥ 18.

Usage

Wrap any agent:

acomplete claude     # Claude Code
acomplete codex      # Codex
acomplete <anything> # any CLI

Type a prompt, pause briefly, and a grey suggestion appears. Tab to accept.

Key	Action
Tab	accept the whole suggestion
→ (Right arrow)	accept one word
Ctrl+R	rephrase the whole prompt in place
Shift+Backspace	undo the last accept / rephrase
keep typing	dismiss / refine

Configuration

acomplete config                       # show current config
acomplete config <key>=<value> ...     # set values

Key	Values	Default
mode	auto · free · byo	auto
provider	anthropic · deepseek · openai · gemini	anthropic
apiKey	your provider key (for byo)	–
model	model id (blank = a fast default)	–
debounceMs	ms of idle before suggesting	400

Auth modes

• auto (default) — reuses the Claude Code login token in ~/.claude/.credentials.json to call a fast model. Zero setup.
• byo — your own provider key:

  acomplete config mode=byo provider=deepseek apiKey=sk-... model=deepseek-chat

• free — a hosted, rate-limited proxy (set proxyUrl). (optional / self-host)

How it works

your keyboard ──▶ acomplete (PTY wrapper) ──▶ agent (claude / codex)
                     │  mirrors your prompt buffer from keystrokes
                     │  debounced call to a fast model for a completion
                     ▼
              grey ghost text drawn at the cursor; Tab injects it

It mirrors your input from raw keystrokes (decoding Windows win32-input-mode records), so it never has to parse the agent's screen. Suggestions come from a separate fast model call — the agent's own loop is too slow for live completion.

Limitations

• Shift+Backspace undo needs a terminal that reports key modifiers (Windows win32-input-mode, kitty, or xterm modifyOtherKeys). On terminals that don't, it falls back to a plain backspace.
• The ghost is truncated to the current row rather than wrapping, so it never collides with the agent's own footer.
• auto mode reuses the Claude Code login token — convenient but undocumented; it falls back silently if the token is missing or rejected.
• Bracketed-paste and Ctrl-modified characters are best-effort in the buffer mirror.

Roadmap

• Self-hostable free-tier proxy
• Demo (see docs/demo.svg)

Contributing

Issues and PRs welcome. Run the tests:

npm test

Debug log (writes to ~/.acomplete/debug.log):

ACOMPLETE_DEBUG=1 acomplete claude

License

MIT © dilyorm