Your agents, one screen. The CLIs you already trust sit beside a browser, a vault, a terminal and a phone, not inside another chat box, and local-first by default.
Windows 10/11 (64-bit) · ~360 MB installer · ~2 GB installed · v0.8.2
| Feature | Where |
|---|---|
| Bots | Desktop |
| Diff sheets | Phone |
Predicts a continuous target variable by fitting a linear function y = w·x + b to the data.
Interface shown is illustrative
Bots, /sidechat and phone parity: queued sends, undo checkpoints, diff sheets and read-aloud.
Release notes →Relay is a shell, not a replacement. It runs the CLIs you already use, keeps their sessions, and adds a built-in chat for cloud APIs or a local GGUF model. Pick a different engine mid-thread; the history comes with you.
Bots, /sidechat and phone parity: queued sends, undo checkpoints, diff sheets and read-aloud.
Bots, /sidechat and a fuller phone app.
Anthropic's CLI, with permission prompts and native tools.
Moonshot's coding agent, including its native modes.
The open-source terminal agent, with its own tools and models.
A prompt-only harness for a fast, focused loop.
A terminal agent with live model costs in the picker.
Plan, standard and accept-edits modes, in a real CLI.
Anthropic, OpenAI, OpenRouter, a custom OpenAI-compatible endpoint, or a local llama-server.
Browse Hugging Face, download a model, run it on your machine. No cloud key.
Expose Relay's tools to any harness, and attach gallery or custom MCP servers on demand.
Vault is plain markdown on your disk. The project wiki cites the file, line range and blob SHA behind every claim. Memory carries decisions between chats, and Session Mesh lets agents ask each other a question.
Your notes as plain markdown, with tags, wikilinks, full-text search, tabs, PDFs and an agent that can read and write them.
A generated knowledge base whose freshness is computed from git. A clean repo costs zero model calls to refresh.
A persistent, evidence-backed memory across chats, with a browser you can search, edit, retire and export.
Agents can see, message and spawn other sessions. Sidechat keeps a private tangent out of the main transcript.
Part of ML Lessons History. Next: 2026-09-16 Logistic Regression.
Linear regression is a supervised learning model that predicts a continuous target variable by fitting a linear function y = w·x + b to the data.
Draw the single straight line that passes as close as possible to all the points at once — squared residuals make a point twice as far off count four times as much.
Training chooses w (slope) and b (intercept) to minimize the mean squared error:
Hours studied are the independent variable; the exam score is the dependent one — as study hours increase, scores rise.
Interface shown is illustrative
A conversation can end in a document, a deck, a diagram, an image or a live visualization, generated on your machine and kept as an artifact you can open, present and improve.
DOCX, PPTX, XLSX and PDF from a plan, with a design layer and page-level checks.
Mermaid diagrams that restyle with your theme and export as transparent PNG.
PNG generation with a local diffusion model. No cloud, no key, nothing leaving the machine.
Local read-aloud and dictation, hands-free with barge-in. Nothing leaves the machine.
Typed charts, tables, timelines and task boards, or a sandboxed HTML mini-app through render_widget.
In unattended work, every write-class call stops at an approval card. A hook can say no, and a limit is visible before it becomes a surprise.
A background run never executes a write-class tool. It queues a deterministic approval with the tool, arguments and a summary.
Run your own scripts before and after every tool call. Deny, ask, observe or rewrite the arguments.
Plan, read-only, manual, auto-edit or full-auto, with always-allow rules and an approval card that shows the real file.
Rolling 1h/24h/7d/30d usage, provider plan windows, user caps and a composer warning notch.
Long-stale threads fold themselves in the background on a configurable idle window. No manual cleanup.
Pair once by QR code or Tailscale. Steer a running turn, undo from a checkpoint, inspect a diff, and have the answer read aloud. The models, keys and files stay on your desktop.
Type while a turn runs; send now or remove the chip.
List and restore a turn from the changes card.
Open a colorized unified diff from a file-edit row.
Speak assistant bubbles and memory rows on device.
Pairing uses a nonce-bound proof and a replay guard.
The phone never receives your provider keys.
Relay Mobile / Android · 0.8.2
One installer, one window. Relay bundles Python and LibreOffice, so the download is larger than a typical editor and the installed footprint is around 2 GB.
Windows 10/11 (64-bit) · ~360 MB installer · ~2 GB installed
Windows + Android · v0.8.2 · updated from GitHub Releases
A local-first desktop workspace for AI coding agents. It hosts the agent CLIs you already use, plus a built-in chat, a terminal, a browser your agent can drive, git tools, a markdown vault and project wiki, persistent bots and memory, and scheduled automations.
No. Bring the CLIs you already run, or use built-in chat with your own model provider. Sessions keep their history, and you can move a conversation between engines mid-thread.
Yes. Browse and download GGUF models from Hugging Face and run them with llama-server. Local image generation, read-aloud and dictation are built in too. Performance depends on your hardware, and local models need additional disk and memory.
Yes. The Android APK pairs with the desktop over a QR code or Tailscale. Queued sends, undo checkpoints, diff sheets and read-aloud are in 0.8.2. The connection is end-to-end encrypted and your provider keys stay on the desktop.
Your workspace, chats, notes and memory live on your machine, in a local database and folders you choose. Cloud providers and connected services still receive whatever you send them. There is no Relay-side account and no telemetry.
Relay checks a signed update feed on startup and every four hours. When a new version is available, the update banner shows the release notes and installs with one click.
Windows 10/11 (64-bit) for the desktop app. The installer is ~360 MB and the installed footprint is around 2 GB because Python and LibreOffice are bundled. Local models and image generation need additional disk, RAM and ideally a GPU.
Relay is proprietary software. The Windows installer and Android APK are free to download from GitHub Releases; the source is not published.
relay.