isair/jarvis: a local voice assistant that keeps a rolling context and redacts before it writes

The problem Jarvis targets: assistants that forget the sentence before yours

Most desktop voice assistants treat each utterance as a fresh command. You say a wake word, you get an answer, and the next question starts from nothing. Jarvis is built around the opposite assumption. The README describes it as "a third person in the room": you can say its name anywhere in a sentence, and it keeps what the README calls a short temporary rolling context of the ongoing discussion. The example given is a group conversation about a picnic where someone asks "Jarvis, what do you think?" and Jarvis resolves that into a weather question using the preceding turns. The target user is someone who talks through problems out loud, wants a dictation hotkey that works in any app, and is unwilling to send audio or transcripts to a hosted service. The README states the privacy position directly: 100% local processing, no subscriptions, no data harvesting, and automatic redaction of sensitive information before anything is saved to disk.

What the pipeline actually does, from wake word to memory write

The example transcripts in the README expose the stages. Speech is transcribed and printed as a Heard line. A separate intent step labelled "Intent (wake word)" decides whether the utterance was directed at Jarvis and rewrites it into a clean instruction, so "Give me a random topic to discuss Jarvis" becomes "give me a random topic to discuss". Tool calls are then selected and shown one per line, for example getWeather, searchMemory, webSearch, or fetchMeals. Only after that does the response get generated. Two details are worth noting because they are design decisions rather than incidental output. First, the rolling context is explicitly short and temporary, which is what allows the assistant to answer "what do you think" without carrying an unbounded transcript. Second, the README claims support for unlimited MCPs and tools "without context rot", and the mechanism it names is smart tool selection: rather than injecting every tool description into the prompt, Jarvis picks a subset per turn. The README also shows echo detection, where Jarvis transcribes its own speech, marks the line as an echo, and returns to wake word mode instead of responding to itself. That is a real failure mode in always-listening systems, and the README treats it as a handled case.

Install paths and the provider switch you have to make

The README offers prebuilt binaries per platform. On Windows you download Jarvis-Windows-x64.zip, extract it, and run Jarvis.exe. On macOS you download Jarvis-macOS-arm64.zip, move it to Applications, right-click, and choose Open. On Linux you download Jarvis-Linux-x64.tar.gz, run tar -xzf on it, and execute ./Jarvis/Jarvis. The only stated prerequisite is a model server: Ollama by default, or any OpenAI-compatible server you already run, with LM Studio, Jan, llama.cpp, vLLM, oMLX and LocalAI named as examples. Provider selection lives under the Configuration section of the README, keyed as LLM Provider. A setup wizard runs on first launch and walks through an initial check, model selection, Whisper configuration, and dictation setup, based on the screenshot filenames in docs/img. The README says Jarvis starts listening automatically once running. One practical consequence of the architecture: because the model server is external, the assistant's latency and memory ceiling are set by whatever you point it at, not by Jarvis itself.

The limitations the maintainers list, and what they imply

The README has a Known limitations block, and it is unusually specific. Jarvis is voice-only, with no text chat interface, tracked as issue #35. There are no mobile apps, issue #17. Stop commands issued during speech are sometimes filtered as echo, issue #24. Dictation does not work on macOS 26 (Tahoe) because of a pynput incompatibility, issue #172. Primary development happens on macOS, and Windows and Linux support may lag. Taken together, these constraints rule out a few use cases outright. If you want to type to your assistant, or read a long answer rather than hear it, this is the wrong tool until #35 is closed. If you are on Linux as a daily driver, you are on the trailing edge of a macOS-first project. And if you are on macOS 26 or later and the dictation hotkey is your main reason for installing, that feature is currently unavailable to you. The project also labels itself as under active development, which is consistent with the release cadence.

Where the licence question sits, and what the release history suggests about upkeep

The repository metadata reports the licence as NOASSERTION, which means GitHub could not classify the licence file into a known SPDX identifier. That is not the same as having no licence, and it is not the same as a permissive one. It means you cannot tell from the metadata alone what you are allowed to do with the code, so read the licence file in the repository root before you build anything on top of it. I am not giving legal advice here, only flagging that the identifier is unresolved. On maintenance, the material shows three releases in the v2.2 line: v2.2.0 on 2026-07-26, v2.2.1 on 2026-08-03, and v2.2.2 on 2026-08-05, with the last push to main on 2026-08-25. That is a tight patch cadence across a single minor version, which fits the maintainers' own description of building in the open. The README also points to an EVALS.md file for tracked accuracy, which is the right place to look if you want numbers, because the README itself does not state accuracy figures.

How this differs from a cloud dictation tool like WisprFlow

The README makes the comparison itself, describing dictation as "like WisprFlow, but free, offline, and private". The difference is not just price. WisprFlow is a hosted transcription service, so audio leaves your machine and the model improves on aggregate data. Jarvis runs Whisper locally through the setup wizard, so the transcription model is a file on your disk and the audio does not travel. That trade has costs the README does not spell out. Local Whisper inference competes for the same GPU or CPU as the language model, and the quality ceiling is whatever model you chose to download rather than a vendor-updated endpoint. The same split applies to the assistant as a whole: a cloud assistant can route to a larger model per request, while Jarvis is bounded by the OpenAI-compatible server you configure. For a user whose main objection is data leaving the machine, that is the point. For a user who wants the best transcription available regardless of where it runs, this is a downgrade they are choosing deliberately.

Who should install this, and what to check before doing so

Jarvis fits a macOS user who already runs Ollama or another local OpenAI-compatible server, wants a wake-word assistant that can follow a conversation across turns, and wants an offline dictation hotkey as a second benefit. It does not fit someone who needs a text interface, a phone client, or a Linux-first workflow, and it does not fit anyone on macOS 26 or later who needs dictation today. Three things to verify before you commit: the licence file in the repository root, since the metadata says NOASSERTION; the EVALS.md file, since that is where the project puts its accuracy tracking rather than in the README; and your own model server's throughput, because Jarvis adds wake-word detection, intent classification, tool selection and response generation on top of transcription, and every one of those stages depends on the provider you configure under LLM Provider. If your local server cannot serve those calls at conversational latency, the rolling context will not feel like a third person in the room.

Editorial conclusion

Adopt Jarvis if you work primarily on macOS, already run Ollama or another OpenAI-compatible server locally, and want a dictation hotkey plus a wake-word assistant that keeps conversational context without sending audio or transcripts off the machine. Do not adopt it if you need a text chat window, a mobile client, or a Linux desktop as your main environment, since the README states that primary development happens on macOS and that Windows and Linux support may lag. Before installing, check the licence file in the repository root, because the GitHub API reports NOASSERTION rather than a named licence, and confirm your macOS version if dictation matters to you: the README lists dictation as unavailable on macOS 26 and later because of a pynput incompatibility.

Official sources

1. isair/jarvis on GitHub
2. Issues
3. README
4. Releases