A coworker that runs on your computer.

Voice in and out, four models on one task, image and audio understanding, a plan that has to prove itself, and a decision that stays yours. Each of these runs on this machine.

• Voice in

  Voice chat — speak instead of type

  Tap the microphone and talk. The clip is recorded in the app, resampled to 16 kHz mono WAV, and transcribed here by LFM3-Audio-2B. The text lands in the composer for you to edit before you send it.

• Voice out

  It answers out loud

  Speak reads any reply aloud, and the header Voice replies switch turns it on for the whole conversation. llama-tts renders the audio with the same LFM3-Audio-2B weight that listens; if that pair is incomplete, the offline Windows system voice takes over and the reply is labelled system-sapi so you can tell which one spoke.

• Hearing

  It listens to recordings in the folder

  transcribe opens a WAV, MP3, FLAC, OGG, or M4A and returns the verbatim text with the measured duration, sample rate, and channel count — never a guess from the file size.

• Vision

  It looks at images

  +img attaches a PNG, JPEG, WebP, or GIF into the workspace, and look reads it through the local vision model. The answer comes back with measured facts: format, byte length, and pixel dimensions.

• Multimodel

  Four models finish one task

  d1-2B decides the shape of the request, LFM3-1.7B acts and writes, LFM3-VL-2B looks at images, and LFM3-Audio-2B hears and speaks. The transcript shows the chosen route, its confidence, and which roles actually ran for that turn.

• Folder

  It reads the folder you chose

  list_files and read_file map and open the real files; search_files matches names and text across the whole workspace; related_files compares one file's distinctive terms against the rest. No shell tool, and no path outside the folder.

• Preflight

  A second model checks the work before you see it

  Before an approval card appears, d1 judges the proposed change against your own words and holds a confident mismatch at a 75% floor. After read-only work, an evidence gate flags thin observations at 65% instead of letting the model guess. Either gate advises or vetoes — it can never authorize.

• Approval

  Every write stays your decision

  A write or edit becomes a card with the exact diff and Allow once / Deny. Pending cards expire after ten minutes, decisions are stored with the transcript, and a path that escapes the folder or passes through a link is refused before it runs.

• Agent loop

  It plans, acts, and proves the result

  The model keeps a live todo list with a check on every step and runs up to twelve tool rounds per turn; Stop cancels the stream and any pending approval. finish is refused until the required output really exists on disk — the report carries the measured byte length and a content digest recomputed from the file it re-read, not a claim.

• Verified

  Verified tasks, when the result must be checked

  The plan-first path freezes acceptance checks, has each step approved before it runs, and retries with a different strategy until the checks pass, keeping a durable journal as evidence. A chat message can never satisfy a mission check.

• Models

  You own the weights

  The Models menu reports each role as Loaded, Installed, or Missing. A download starts only when you click it, runs one role at a time with live progress and cancel, and is checked against the published byte length and SHA-256 before it is used or hot-swapped into an idle chat.

• Offline

  Local by construction

  All four servers bind to 127.0.0.1. No account, no API key. Under LFM3_PROFILE=packaged the engine refuses to start with a non-loopback endpoint or with the download flag on. With tracing left off, the only network traffic is a model download you start or a link you open.

• Sessions

  Conversations survive restarts

  Chat history, pending approvals, and jobs live in SQLite inside your data folder. Rename, delete, or resume any session; a turn the app stopped mid-way comes back marked cancelled rather than being replayed.

• Capacity

  32,000 tokens, on a CPU, on both desktops

  The acting model is served with a 32,000-token window, so long sessions keep their history. The bundled llama.cpp runtime is CPU-only — no GPU required or configured — and the same app ships for Windows 10/11 and Linux x86_64, with a headless lfm3 CLI for Linux.

The speech round trip (spoken answer recorded back through the local audio model) and the decision gates are measured by the project's own multi-model benchmark against the pinned weights. The README publishes those numbers with their caveats rather than rounding them up.

1. d1-2B

   Route the request

   A compact decision model classifies what you asked for and picks the roles that turn needs. Small talk skips it. Routing can widen capability; it can never grant permission.

2. LFM3-1.7B

   Act on it

   The acting model plans in short steps, calls file tools, reads the real output, and writes the deliverable. It is served with a 32,000-token window, so long sessions keep their history.

3. TOOLS

   Touch only your folder

   List, search, read, compare, write, edit. Every path is checked against the workspace you chose; anything outside it is refused before it runs. Chat has no shell.

4. YOU

   Approve the change

   Reads and searches run immediately. A write or edit becomes a card with the exact diff and Allow once / Deny. A denied change leaves the file untouched.

All model servers bind to 127.0.0.1. With tracing off (the default), the only network traffic is a model download you start, or a link you open.

Version 0.1.1 · released 1 November 2026. Every package includes the pinned llama.cpp runtime and no model weights. All four files are attached to the v0.1.1 release, with the changes since v0.1.0, install steps, and checksums.

1. Install. Run the Windows setup, install the Linux .deb or run the .AppImage, or unpack the headless CLI archive.
2. Download models. Open Models and fetch the roles you need.
3. Pick a folder. Ask for the work, then approve the diff.