opencues/opencues

opencues

Turn any text field into a two-way LLM channel: it reads what you write and fills what you ask. Flags a slip as you type, answers anything you end with _. Claude Code, OpenCode, Gemini CLI, shell, Chrome, DeepSeek Harness. Model-agnostic, open standard, no chat window.

待分类agentic-aiaiai-agents

#

Turn any text field into a two-way LLM channel. It reads what you write and fills what you ask: catching a slip as you type, or answering the moment you end a line with _. A drop-in for Claude Code, OpenCode, Gemini CLI, your shell, and Chrome. Model-agnostic, output you review before it sends, an open standard with no chat window.

The model comes to your cursor, both ways. Cues react to what you've already written and surface a fix or a sharper line inline, unprompted. Blanks act on demand: end a line with _ and the model fills in the rest. No chat window, no copy-paste, no context switch.

OpenCues is platform, model, and provider agnostic, engineered from the ground up to enable native inline AI.

You type · You get i keep having to approve every single git command · ↳ 💡 Tired of approving? /permissions allow rules like Bash(git *), or Shift+Tab to auto mode ok this is a mess, let's start over on the auth stuff · ↳ 💡 Starting over? /clear wipes the conversation, CLAUDE.md stays let's ship it Thursday the 19th · ↳ ⚠ the 19th is a Friday we should probably go ahead and refactor this · ↳ we should refactor this hey can u send me that report when u get a sec make this formal _ · Could you please send me that report at your earliest convenience? hello world translate to japanese _ · こんにちは世界 draft an email to my landlord asking for a rent reduction _ · (the email, written) ffmpeg command to convert a video to web-ready mp4 _ · ffmpeg -i input.mov -vcodec libx264 -crf 23 -pix_fmt yuv420p -acodec aac output.mp4

Rows with _ are blanks: you ask, the model fills in. Rows without are cues: the model speaks up on what you wrote, no prompt. A 💡 cue knows the situation you are in and _ makes the draft the command.

#

Quickstart

npm install -g opencues              # needs Node 22+ and git
opencues set-key cerebras csk-...    # cerebras.ai — free tier, lowest latency
opencues install claude-code         # or: opencode | gemini-cli | chrome | shell | dsh
claude-cues                          # launch — native `claude` is untouched

Full walkthrough, prerequisites, and per-host detail: docs/install.md. opencues doctor diagnoses anything that looks wrong.

#

Integrations

Host · Status · Install Claude Code · Available · opencues install claude-code OpenCode · Available · opencues install opencode Gemini CLI · Beta · opencues install gemini-cli Chrome · Beta · opencues install chrome Shell · Beta · opencues install shell DeepSeek Harness · Beta · dsh plugin --profile web add @opencues/dsh

Each pins its own upstream fork and never touches your native host install. DeepSeek Harness is the exception: it has a real plugin system, so there is no fork to pin and no OpenCues CLI step — see integrations/dsh/README.md.

⚠️ Windows: not supported natively, run inside WSL2 (see docs/install.md).

#

What you get

Feature · What it does Blanks · Type _ for free-form generation, translation, formatting, full rewrites, or keyword-bound system actions (volume _, weather _). Sentence rewrites · Cycle a whole sentence to a different register (formal, concise, ...) seamlessly, no _ needed. Word cues · Navigate to a single word and cycle a smaller LLM-suggested alternative. Live actuators · volume _ reads the real level and leaves a knob you can turn: Ctrl+Alt+↑/↓ or a bare _ moves it, and the device follows. Dismissing a cue · A cue that only tells you something can be silenced from its own note: _ once quiets it for a while, _ again forgets it for good. opencues dismissals lists what you forgot and turns any of it back on. Session-contradiction cues (opt-in) · Flags a draft that goes against a decision made earlier in the same coding session, on any host with a transcript. Ask-cues (opt-in) · Turns a vague sentence into an inline question with cyclable answers, using the assistant's own clarifying prompt. Personal + ambient context (opt-in) · my email _ substitutes your real address; fluid lookups can read the page you're on. Hot-reload · Every .md config picks up edits in ~2s, no restart.

Full feature catalogue (44 concepts): docs/features/README.md. Running a lot of these at once? opencues usage prices every LLM call your hosts have made, across features, so the total isn't a guess.

Configuration & LLM providers

Config lives at ~/.cues/ — one OPENCUES.md for runtime settings, plus per-surface source folders. Seven providers supported; set an env key or opencues set-key and you're done.

Full reference: docs/configuration.md · docs/guides/llm-providers.md (switching provider/model, free mode, failover).

#

Security

OpenCues has no tool handlers or exec layer for LLM output — no MCP-tool execution, no agentic actions, no side-effect channel. Worst-case, an LLM response lands as user-visible text in the buffer you review before submitting. That single invariant is what keeps prompt injection a UX failure instead of a data-exfiltration channel, across every surface below.

23 of 27 tracked attack classes closed, 3 closed-with-caveat, 1 tracked for the future pack registry (full audit table).

Defense · What it covers Sandbox isolation · Third-party blank JS runs in a real V8 isolate (isolated-vm) — its own realm, own intrinsics, no sandbox-escape via constructor-chain pivots Capability gates · A blank only gets network/llm/storage/secrets access if declared; secrets without a matching host binding are refused at load time Resource quotas · Sliding-window caps on fetches, LLM calls, and storage writes — no polling hammer, no runaway LLM burn Output sanitization · Blank output is stripped of HTML/script tags, zero-width chars, and bidi overrides before it reaches the buffer

Check a pack before trusting it:

opencues review ./untrusted-pack/

Full threat model: docs/architecture/security-audit.md. Reporting a vulnerability: SECURITY.md.

Contributing

What · Where New host integration · docs/guides/adding-an-integration.md New cue/blank · docs/guides/adding-a-cue-blank.md Working on the reference runtime · CONTRIBUTING.md New to the terms · docs/glossary.md

Join the community — questions, feedback, and the people building alongside you. We're also part of OpenSourceIRL, a community for people building in the open.

#

License

Apache License 2.0.

#