Radiant brings every model, a library of expert agents, real tools, and a live terminal into one Mac-native workspace, and keeps every decision in your hands.
Radiant gives the model a real project to work in, shows you the plan, asks before it acts, and records every move. You stay the operator, from intent to verified change.
Start from 142 ready-made specialists across two dozen fields, then tune each one’s model, personality, and skills. Duplicate an agent, export a curated pack, or bring several into one group chat and let them build on each other.
Sign in with your Claude, ChatGPT, Grok, Qwen, or Copilot subscription, no API key needed, or paste a key for any provider. Run Ollama and LM Studio models with none at all, and download GGUFs from Hugging Face with a disk-space check before every pull. Switch models mid-thread and your whole session comes with you.
Radiant keeps the agent moving and gives you the right checkpoint at the right moment. Nothing happens off-screen.
Attach files, point at a workspace, pick an agent, or start from a saved recipe.
Plan mode researches first, asks the questions that are genuinely yours, and waits before touching anything.
Read and edit files, run shell commands, keep dev servers alive, and drive a real terminal or the browser.
A live activity feed, changed-file chips, command output, and test results close the loop.
The power features stay discoverable without turning the workspace into a control panel.
Ask for every command, auto-run the low-risk ones, or allow all, so a long build finishes without stopping to knock. Add a sandbox macOS enforces, and commands can only change files inside the project, with the internet off if you like.
Agents render stat tiles, tables, diffs, and clickable choices inline.
Let a vision model drive the browser and desktop, opt-in and visible. On the desktop it asks about each app by name first, and terminals are never allowed.
Capture its HTML, CSS, and a screenshot straight into the chat.
Watch a page’s network calls and build a reusable HTTP client from them.
Add reusable skills globally or per agent, connect MCP tool servers, and let Radiant suggest skills from repeated work.
A kanban board where the run moves the card — Working, Needs you the moment it asks, Review when it’s done. Each task keeps its comments, runs and worker logs in one feed, can wait on other tasks, and can be split into subtasks by the agent itself.
A loop is a run of steps, each with a check it has to pass — a command that must succeed or a sentence another agent judges — and a step that fails goes round again with the reason. A graph runs the steps that don’t depend on each other at the same time, shows what each round cost, and stops at a budget you set. A check that sees a wall says so and ends the run instead of retrying.
A panel of every file the agent changed, with an inline diff, a commit box for the files you pick, and Undo that stays undoable for an hour. Nothing is ever pushed.
Continue the sessions you started in either app, or pick one as a chat’s model. They use their own sign-in, so there’s no API key. Opt-in, and private to your Mac.
Open chats sit in tabs, long runs fold to “Worked for 2m 39s,” and the model, thinking level and permission mode are always on screen in words.
Every reply that edits files leaves a checkpoint. Roll the files back to before it, and keep the conversation.
An easy message can go to a fast model from the same provider, decided in a third of a second. Or choose Jev Router on OpenRouter and let it pick from every model it can reach. Each reply is labeled with the model that actually answered.
No cap on how many steps a job takes. The only ceiling is a spend budget you set, measured in real cost, and a turn that keeps failing the same way is stopped instead of burning tokens. As a chat grows, old tool results and earlier side conversations the task no longer needs are set aside instead of summarized away, and what it still depends on keeps its place.
Radiant reads a project’s AGENTS.md, CLAUDE.md, .clinerules or .cursorrules into every chat, the way other coding agents do, and shows the branch’s pull request and CI status above the box you type in.
Hold a spoken conversation over any chat, on OpenAI GPT-Live or Google Gemini Live. The voice only listens and speaks — your own model, tools and approvals do the work, and the transcript is saved into the chat.
The iPhone and iPad app runs open models directly on the device with Apple’s MLX — search Hugging Face, see whether a model fits before you download it, and chat with nothing leaving the device.
If a model errors, runs out of context or the connection drops, the reason is written into the conversation with a Continue button — never an empty reply. The token counter shows how much was served from cache, so a busy chat can be told from an expensive one.
Isolate an agent on a private Linux desktop in a container, never touching your Mac.
Run Radiant on the Mac that stays awake and reach the same agents, models, and sessions from your other Macs.
A dozen palettes, per-agent color, ten motion backgrounds, and ⌘K to switch models or jump sessions.
Your API keys live in a locked config on your Mac and never enter the browser. Local models need no key at all, the server binds to your machine, and every command and file change shows up in the activity trail before and after it happens.
Radiant for Apple Silicon Mac. Free to download, and MIT licensed — the whole thing is on GitHub.
There is an iPhone and iPad app too, on the App Store. It runs open models directly on the device — nothing you type has to leave it.
