On this page
  • Chat with an agent and see what it does: its reasoning, the tools it calls, and what they return.
  • Use the model you choose: a Claude, ChatGPT, or SuperGrok subscription, an API key, or a model running on your computer.
  • Approve or reject actions before an agent takes them.
  • Run many agents at once and watch them on a map.
  • Run agents on this computer and on other computers you connect to, from one window.

Download

Version 0.8.910 November 2026Release notes

PlatformFileSizeSHA-256
Windowsx64 installerADF-Studio-Setup-0.8.9.exe170 MBca8db69dc7c6…
macOS (Apple Silicon)M1 and laterADF-Studio-0.8.9-arm64.dmg227 MB9cfb12ff1f3f…
macOS (Universal)Intel and Apple SiliconADF-Studio-0.8.9-universal.dmg329 MBe83202e2db95…
Linux (Debian, Ubuntu)x64adf-studio_0.8.9_amd64.deb166 MBe302b734c64f…
Linux (AppImage)x64, needs libfuse2ADF-Studio-0.8.9.AppImage230 MBf596ca154ebb…

Studio is an early release. The format and app may change between versions. Older builds are on GitHub Releases.

Documentation

Screenshots

The ADF Studio window: agents in the sidebar, a Markdown document in the editor, and the Loop panel showing a conversation with tool calls, reasoning blocks, and token counts.
Fig. 1. The main window: agent list, document editor, and the Loop.
The Mesh view: two agent districts on a dark hex-tile map with status chips, a terrain legend, and a minimap.
Fig. 2. Mesh view.
An approval modal: an agent requests a sys_update_config call, with formatted arguments and Approve and Reject buttons.
Fig. 3. Tool approval.
The Loop tab: a context block, a user message, an expandable thinking block with a token count, assistant text, and inline tool-call chips.
Fig. 4. Loop transcript.
The Agent config panel: Identity fields and Model fields including provider, model ID, temperature, max tokens, and reasoning effort.
Fig. 5. Agent configuration.
The Home dashboard: status tiles for providers, MCP servers, channels, packages, containers, agents, and networking.
Fig. 6. Home dashboard.
Settings, Providers: connected providers listed with logo, name, and default model.
Fig. 7. Providers.

Features

FeatureWhat it does
LoopChat with an agent. The transcript shows your messages, the agent’s reasoning, tool calls, results, and token counts.
Agent configurationIdentity, model, instructions, tools, triggers, messaging, serving, and timers, plus the raw config.
FilesThe agent’s files: README.md, mind.md, soul.md, and any uploaded or agent-written files.
ProvidersClaude, ChatGPT, and SuperGrok subscriptions (Claude through the Claude Code sign-in on the same computer), hosted APIs (Anthropic, OpenAI, OpenRouter, Gemini, and others), and local servers such as LM Studio or Ollama. Other OpenAI-compatible chat completions servers work when both the server and model support native tool calling.
ConnectionsAgents run in the local server (the ADF daemon), shared with the terminal app. Saved connections add the servers of other computers, each listed as a runtime of its own, and Share lets other devices use this one.
Mesh viewEvery agent as a tile on a hex map, grouped by directory, with other runtimes as platforms at the edge. Select and command groups with hotkeys.
ApprovalsRestricted tools pause for human approval, inline on the Mesh view or in a full-context modal.
Home dashboardStatus tiles for providers, MCP servers, channels, packages, containers, agents, and networking.
MCP serversRegister tool servers and their credentials; attach them to agents.
Code executionSandboxed code with a shared package store available to every agent on the instance.
ComputerIsolated agents can get a small Linux desktop with a managed Chromium, streamed into Studio.
NetworkingAgent-to-agent messaging over ALF, LAN discovery (mDNS), and HTTP serving of agent content.
MultimodalImage, audio, and video input when the model and agent config enable it.
UpdatesChecks for new releases hourly and on demand from Settings → About.

System requirements

OSCPUPackageNotes
Windowsx64.exe (NSIS installer)
macOSApple Silicon.dmg (arm64)Signed and notarized
macOSIntel or Apple Silicon.dmg (universal)Signed and notarized
Linuxx64.debsudo apt install ./adf-studio_*_amd64.deb
Linuxx64.AppImageNeeds libfuse2
  • At least one model provider: an API key, a local model server, a ChatGPT or SuperGrok subscription, or a Claude Pro or Max plan with Claude Code installed.
  • Local models need native tool calling (function calling) for agent tools. Vision is optional: understanding images or screenshots needs image support in the server and model, with image input enabled for the agent. See Local models for server setup requirements.
  • Optional: Podman for container-backed compute.
  • On first launch Studio creates ~/.adf-studio/ for settings, the shared package store, and compute state.

Build from source

Building current ADF main requires Node.js ^22.23.4 || ^24.22.1 || ^26.12.1 and npm. The packaged desktop app includes its own runtime.

git clone https://github.com/christianbalevski/adf.git
cd adf
npm install
npm run dev        # launches ADF Studio

npm run package builds platform installers with electron-builder. See README and RELEASING.md.