Key Takeaways
- Parlor (github.com/fikrikarim/parlor) is a free, open-source, on-device real-time multimodal AI assistant β a research preview, not a finished product
- Created by developer Fikri Karim, who also built a separate hosted voice-AI project called Bule AI; Parlor grew out of wanting a fully local alternative after trying OpenAI's GPT-Live
- Apache 2.0 licensed, confirmed via the GitHub repository's LICENSE file
- Runs Gemma 4 (Google DeepMind), specifically the E4B variant by default, through llama.cpp for combined speech and vision understanding
- Over 2,000 GitHub stars and 271 forks as of this review
- Two tagged releases so far: v1.0.0 (July 29, 2026) and v2.0.0 (August 2, 2026)
π In One Sentence
Parlor is a free, open-source, on-device real-time voice-and-vision AI assistant built by a solo developer as a research preview, running Gemma 4 through llama.cpp on a Mac with Apple Silicon or a Linux machine with a supported GPU, with over 2,000 GitHub stars.
π¬ In Plain Terms
Parlor is software you install and run yourself that lets you talk to an AI assistant out loud, with your webcam optionally watching too, entirely on your own computer by default β no cloud voice subscription, no account, and no per-minute cost, unless you opt into an optional cloud research feature. It needs a fairly capable Mac or Linux machine with a GPU, since it runs the AI model locally.
πNote: This review is based on Parlor's own GitHub repository, its README, and its release history via the GitHub API. It does not claim PromptQuorum has independently reproduced Parlor's published latency benchmarks β those figures are the project's own, measured on an Apple M3 Pro.
What Is Parlor?
Parlor is a free, open-source research preview of a fully on-device, real-time AI assistant that listens, watches through a camera, and speaks back β comparable in concept to OpenAI's GPT-Live, but running locally instead of in the cloud. Its README explicitly labels it a "research preview" and warns to expect rough edges and bugs.
- Product type: a self-hosted, local web application β you run a server on your own machine and open it in a browser
- Creator: developer Fikri Karim, working solo; the README states the codebase was written "with strong assistance from Claude and with humans leading the ideas, testing, and debugging"
- Repository: github.com/fikrikarim/parlor, created April 5, 2026
- License: Apache 2.0, confirmed via the repository's LICENSE file
- Funding: no funding round, investor, or commercial backing was found for this review β treat Parlor as an independent, solo-developer research project rather than a funded company
- Scale: over 2,000 GitHub stars, 271 forks, as of this review
How Does Parlor Actually Work?
Parlor uses a cascade pipeline, not a single end-to-end voice model: your browser streams microphone audio and camera frames over a WebSocket to a local FastAPI server, which runs turn-detection, a vision-language model, and text-to-speech as separate stages.
- Turn detection: browser-side Silero VAD flags silence, then Pipecat's smart-turn-v3 model (about 20ms) judges whether you actually finished speaking, so mid-sentence pauses are not mistaken for the end of your turn
- Understanding and generation: Gemma 4 (Google DeepMind), the E4B variant by default, running through llama.cpp with QAT 4-bit quantization, processes both the audio and camera frames and streams a reply
- Action handling: a separate grammar-forced JSON request from the same model decides on timers, mode switches, or research requests, so control instructions never leak into the spoken reply
- Speech output: Kokoro text-to-speech (MLX backend on macOS, ONNX on Linux) speaks the reply sentence-by-sentence while the model is still generating
- Barge-in: you can speak over Parlor to interrupt it β generation is aborted server-side rather than just muting playback
- Optional cloud research: if you set a
REASONER_API_KEY, Parlor can hand off open-ended research questions (e.g. "find the best pizza in Rome right now") to a frontier model over an OpenAI-compatible endpoint such as OpenRouter, while the local conversation continues; without that key set, Parlor stays fully on-device
What Can You Do With Parlor?
Parlor supports hands-free spoken conversation with optional camera input, plus a handful of named modes for specific tasks.
- Hands-free conversation: no push-to-talk button β Parlor detects when you start and finish speaking on its own
- Camera-aware answers: point your webcam at something and ask about it; frames are streamed alongside your speech to the same local model
- Timers: "set a timer for three minutes for the pasta" β the server itself owns the countdown clock (not the language model), so the timer still rings even during a silent turn, and a cancel-able countdown chip tracks it on screen
- Live translation mode: say "translate everything I say into English" and Parlor becomes a consecutive interpreter, rendering each utterance after a short pause, in any language Gemma 4 understands, until you say "stop translating"
- Just-listen mode: say "just listen for a while, I want to think out loud" and Parlor transcribes everything on screen without speaking back, until you address it again
- Background research: with an optional cloud API key configured, Parlor can delegate open-ended lookup questions to a separate model while the local conversation keeps going, then speak the answer once it arrives
Platform, Pricing, and Licensing
Platform
- What Parlor states:
- A self-hosted local web app, accessed through a browser. Runs on macOS with Apple Silicon, or Linux with a supported GPU. No Windows build exists.
Cost
- What Parlor states:
- Free and open source, no subscription, no account. All AI processing is local by default; an optional cloud research feature requires your own API key for a separate provider.
Licensing
- What Parlor states:
- Apache 2.0, confirmed via the GitHub repository's LICENSE file.
Hardware
- What Parlor states:
- About 6 GB of free RAM for the default Gemma 4 E4B model; the smaller E2B variant fits in about 4 GB; a larger 12B option needs about 8 GB.
Status
- What Parlor states:
- Explicitly labeled a "research preview" in its own README β expect rough edges and bugs, per the project itself.
Verify current hardware and platform requirements directly on github.com/fikrikarim/parlor before installing, since a research preview can change these between releases.
Install Parlor
Parlor installs from source via the uv Python package manager; there is no packaged installer for any platform.
GitHub repository (source code, Apache 2.0)
Requirements
- Link:
- Python 3.12+, llama.cpp (build b9503 or newer), macOS with Apple Silicon or Linux with a supported GPU
Install command
- Link:
git clone https://github.com/fikrikarim/parlor.git, thenuv syncanduv run parlor
First run
- Link:
- Open
http://localhost:8000, grant camera and microphone access; models (about 5.7 GB) download automatically
Parlor requires a terminal, a Python environment, and a local llama.cpp install (brew install llama.cpp on macOS) β there is no GUI installer.
Parlor vs. Voxa
Parlor and Voxa are both open-source, real-time voice assistants, but Parlor is deliberately local-only while Voxa is hybrid by design.
Processing location
- Parlor:
- On-device only by default; cloud research is opt-in and separate from the conversation itself
- Voxa:
- Hybrid β routes conversations through cloud realtime models (Gemini Live, OpenAI Realtime) or a self-hosted local daemon you configure
Camera input
- Parlor:
- Yes β camera frames feed the same model as your speech, for vision-aware answers
- Voxa:
- Not part of its core feature set per its own documentation
Platform support
- Parlor:
- macOS with Apple Silicon, or Linux with a supported GPU; no Windows build
- Voxa:
- Desktop app built with Tauri v2, cross-platform by that framework's design
Interface
- Parlor:
- Browser-based local web app
- Voxa:
- A frameless, always-on-top orb you tap to start a conversation
Maturity
- Parlor:
- Explicitly labeled "research preview"; two tagged releases as of this review
- Voxa:
- Positioned as a usable desktop assistant rather than a labeled research preview
Both are open-source and free. Choose Parlor if you want camera-aware, fully local-by-default processing on Mac or Linux; choose Voxa if you want a lighter desktop orb with the option to fall back to cloud realtime voice models. See the Voxa review for details.
Who Should Use Parlor?
Parlor fits developers and technically comfortable users on Mac or Linux who want to experiment with a fully local, camera-aware voice assistant and are willing to tolerate research-preview rough edges.
What Parlor Is Not Good For
Parlor is not a good fit if you need Windows support, a one-click install, or a stable production tool rather than an evolving research preview.
- Not for Windows users β only macOS with Apple Silicon or Linux with a supported GPU are supported
- Not a packaged, one-click app β installation requires a terminal, Python 3.12+,
uv, and a workingllama.cppbuild - Not a finished, versioned product β its own README calls it a research preview and warns to expect bugs
- Not multi-user or hosted for others β it is designed to run locally for one person on their own machine
- Not backed by a company or funding round that this review could verify β treat it as an independently maintained, solo-developer project
Common Mistakes When Evaluating Parlor
Most confusion about Parlor comes from expecting a finished commercial product, or underestimating its hardware and platform requirements.
Competitors and Alternatives
Parlor is most directly comparable to other open-source, real-time voice and personal-assistant projects that emphasize local or hybrid processing over a pure cloud subscription.
Tool | Best known for | Link |
|---|---|---|
| Voxa | Open-source desktop voice assistant orb, hybrid local/cloud realtime voice routing | Voxa review |
| Jarvis (Mac) | Mac-focused voice assistant app built around local and configurable AI backends | Jarvis Mac review |
| nanobot | Self-hosted personal AI agent with persistent memory, text-first rather than voice-first | nanobot review |
| OpenAI GPT-Live | Cloud, subscription-based realtime voice product named as Parlor's inspiration | openai.com/index/introducing-gpt-live |
This list reflects tools commonly compared to Parlor in the local-voice-assistant space, not an independent PromptQuorum ranking β verify each tool's current platform support and feature set before choosing.
Frequently Asked Questions
What is Parlor?
Parlor (github.com/fikrikarim/parlor) is a free, open-source (Apache 2.0) research preview of an on-device, real-time voice-and-vision AI assistant, built by a solo developer to run entirely on a Mac or Linux machine.
Is Parlor free?
Yes, Parlor is free and open source. It requires no subscription or account. An optional background-research feature needs your own API key for a separate cloud provider if you choose to enable it.
Does Parlor work on Windows?
No. Parlor requires macOS with Apple Silicon or Linux with a supported GPU. There is no Windows build.
Does Parlor send my voice or camera data to the cloud?
By default, no β microphone audio and camera frames are processed locally through a model running on your own machine. Data only leaves your device if you explicitly set a REASONER_API_KEY to enable the optional background-research feature.
How do I install Parlor?
Clone the GitHub repository, install the uv Python package manager and llama.cpp, then run uv sync followed by uv run parlor. Open http://localhost:8000 in a browser and grant camera and microphone permissions.
What AI model does Parlor use?
Gemma 4 from Google DeepMind, running locally through llama.cpp β the E4B variant by default, with smaller (E2B) and larger (12B) options available depending on your hardware.
Who created Parlor?
Developer Fikri Karim built Parlor as a solo research preview, after previously self-hosting a separate always-online voice AI project called Bule AI.
How much RAM does Parlor need?
About 6 GB of free RAM for the default E4B model configuration; the smaller E2B variant fits in about 4 GB, and the larger 12B option needs about 8 GB.
Can Parlor see through my camera?
Yes. You can point a webcam at something and ask Parlor about it β camera frames are streamed to the same local model alongside your speech.
Is Parlor a finished product?
No. Its own README explicitly labels it a "research preview" and asks users to expect rough edges and bugs. It has two tagged releases so far.