# CueLight (slim)

> CueLight is an open-source Python library and set of open-weight models that estimate, every 20 ms of streaming audio, whether a voice-agent user has finished their turn. It runs locally from audio alone, in English.

- Full: https://www.anchorterminal.com/tools/cuelight.md (~5,500 tokens) · this version ~1,380 tokens · JSON https://www.anchorterminal.com/tools/cuelight.json · canonical https://www.anchorterminal.com/tools/cuelight
- Index: https://www.anchorterminal.com/llms.txt · API: https://www.anchorterminal.com/api/v1/index.json · Updated: 2026-10-15
- Licence: CC BY 4.0, credit Anchor Terminal (anchorterminal.com) and link this page when you pass anything on

**D · 48.8/100 · rank #873 of 1094 · #14 in Conversational voice agents · not agent-ready · confidence medium**


Assessment: CueLight gives a voice pipeline a local end-of-turn probability every 20 ms from audio alone, with a 1.9 MB model and Apache-2.0 weights. It is one day old, English only, installs from Git rather than PyPI, has no CI or changelog, and its accuracy figures are the author's own.

## Facts

- Kind: SDK + MCP · vendor: sshah03 (individual developer, name not stated) · category: Conversational voice agents · legal entity: sshah03 (individual GitHub account, no name or organisation stated) · provenance 40/100
- Auth: None · pricing: Free · x402: no · licence: Apache-2.0
- Probe metrics: not measured yet (probes haven't run)
- Surface listed: The Python package `cuelight` 0.1.0 (`TurnDetector`, `Endpointer`, `EndpointPolicy`, `load`) and its three model repositories on Hugging Face, read from the source, the README and the model cards on 15 October 2026. Nothing was installed or run
- Interface: `TurnDetector.push(chunk)` takes 320 float32 samples (20 ms at 16 kHz mono) and returns p(end of turn), with `p_speech` for the same frame. `Endpointer.push(chunk)` returns True once when the turn ends
- Models: cuelight-nano (0.5M parameters, 1.9 MB, ONNX step file), cuelight-nano-ssm (0.5M, experimental, ONNX step file), cuelight-base (88M, 336 MB, PyTorch only)
- Presets: `snappy` (default, about 1 in 10 pauses interrupted by the author's measure) and `careful` (about 1 in 20), or a custom `EndpointPolicy(threshold, min_silence_s, timeout_s)`
- Languages: English only. The README says other languages are not planned soon
- Runtime: Python 3.10 or later, PyTorch 2.1 or later, numpy, safetensors, huggingface_hub. Weights load from safetensors
- Framework integrations: None in the project, which says it depends on no voice framework. OpenVoiceOS maintains a plugin wrapper, whose README says it is not on PyPI yet
- Telemetry: None found in the package source (searched 2026-10-15). Weight downloads go through huggingface_hub, whose own settings apply
- Project age: One day at the check. One commit, no tags, releases, issues or pull requests
- Releases in 90 days: None tagged. One published version, 0.1.0 in pyproject.toml, on 14 October 2026

- Scores: Reliability 29, Performance pending, Schema & documentation 50, Agent ergonomics 61, Security & auth 50, Payments & pricing 60, Task success pending, Maintenance & community 46, Transparency & trust 54 · total over the 7 assessed categories
- Why: Reliability, Read as local software. · Schema & documentation, Library reading. · Agent ergonomics, Library reading. · Security & auth, Local-software reading. · Payments & pricing, No payment protocol (0). · Maintenance & community, Published on 14 October 2026, one day before the check (30). · Transparency & trust, Apache-2.0 for code and weights, with a NOTICE that lists each training dataset, teacher model and licence, and the conditions passed on fro…
- Sources: 9, open questions: 6, both in the full twin
- Capabilities: voice.pipeline, inference.open-weights
- JSON: https://www.anchorterminal.com/api/v1/tools/cuelight.json
- Verify (for the vendor): the badge `https://www.anchorterminal.com/badges/cuelight.svg` or a link to https://www.anchorterminal.com/tools/cuelight from a page on sshah03.github.io or one of its subdomains, or the README of github.com/sshah03/CueLight, then `POST https://www.anchorterminal.com/api/v1/verify` `{"slug", "url"}` or `verify_listing` at /mcp; re-checked weekly, no effect on the grade. Snippets in the full twin.

## Before you call it

1. Feed 320 float32 samples at 16 kHz mono per `push`. Resample first, or the call raises an AssertionError
2. Spell the preset as `snappy` or `careful`. Any other name silently uses default thresholds
3. Call `clear()` before a new call or session. Pass `reset_on_turn_start=False` if your pipeline already resets per turn
4. Pin weights by downloading them to a local folder and passing that path to `from_pretrained`
5. Don't use it to identify speakers or clone voices. The training data's conditions forbid it

## Connect

```bash
pip install git+https://github.com/sshah03/CueLight
```
Full config and headless snippets are in the full page. Through letme (picks today, calling later): https://letme.dev/cuelight

## Similar tools

| Tool | Grade | Score | Shared capabilities | Slim |
| --- | --- | --- | --- | --- |
| GroqCloud | BB | 75.6 | inference.open-weights | https://www.anchorterminal.com/tools/groq.min.md |
| ElevenLabs Agents API + MCP | BB | 71.3 | voice.pipeline | https://www.anchorterminal.com/tools/elevenlabs-agents.min.md |
| Mistral AI API | BB | 71.2 | inference.open-weights | https://www.anchorterminal.com/tools/mistral-api.min.md |
| Retell AI API + MCP | B | 69.1 | voice.pipeline | https://www.anchorterminal.com/tools/retell-ai.min.md |
| Cloudflare Workers AI | B | 68.3 | inference.open-weights | https://www.anchorterminal.com/tools/cloudflare-workers-ai.min.md |


## Panel reviews (0, desk reviews from public material, no calls made)