Scope

This document defines what belongs in the Desktop-AI-Agents list and what doesn't. It's the source of truth for inclusion decisions — if something is ambiguous, this file should settle it (and if it can't, the file should be updated).

Unifying principle

The agent loop runs on the user's device.

The underlying model may be hosted (Anthropic, OpenAI, xAI, Mistral, Moonshot, GitHub Copilot…) or fully local (Ollama, llama.cpp, vLLM). What matters is that the harness — the thing that plans, calls tools, reads the filesystem, shells out, keeps state — lives on the user's device.

"Desktop" is shorthand for local / on-device

The repo name says "Desktop" but that's shorthand. In practice, in-scope hardware includes:

• Desktops and workstations — the prototypical case
• Laptops — obviously
• Embedded devices and SBCs — Raspberry Pi, mini-PCs, dedicated agent hardware (e.g. mimiclaw-style $5-chip deployments)
• Mobile devices — phones and tablets, if the agent loop actually runs on the device rather than calling out to a hosted backend
• Home servers / NAS — self-hosted boxes on the user's LAN count as "on-device" for the purposes of this list, even though they're technically servers; the distinguishing line is user-owned and user-controlled, not "cloud-hosted SaaS"

The unifying test is still the same: does the agent loop run on hardware the user owns and controls? If yes, it's in scope regardless of form factor.

In scope

• CLI / TUI coding agents you run inside a repo (aider, opencode, goose, OpenHands, copilot-cli, codex, kimi-cli, mistral-vibe, grok-cli…)
• Sysadmin-style agents that fluidly use the filesystem, shell, and networked resources (local or LAN) rather than being constrained to a repo
• Tooling that wraps, manages, or orchestrates on-device agent CLIs (session managers, task managers, parallel orchestrators)
• MCP servers and tool providers that plug into desktop agents
• Multi-agent frameworks whose orchestration runs locally
• Agent operating systems / computer-use agents that drive the host OS
• Protocols and specifications for desktop agent interop (AGENTS.md, Agent Skills, AG-UI, MCP-adjacent specs)
• Harnesses and frameworks for building desktop agents
• Desktop-facing capabilities like wallets, payments, browser automation that are designed to be embedded in an on-device agent

Vendor (first-party) and third-party projects are both in scope and are tagged separately.

Out of scope

• Server-deployed / cloud-hosted agent platforms. If the agent loop runs in someone else's cloud, it's out — even if there's a thin CLI that talks to it.
• Hosted SaaS agents with no meaningful local execution.
• IDE-first tools where the CLI is only a companion to an IDE extension (Cursor, Cline-as-VSCode-extension, Continue…). If the CLI is a first-class way to run the agent, it's in.
• Pure inference engines without an agent loop. Ollama, llama.cpp, vLLM, LM Studio (as a server) are model runners, not agent runners. They're used by things in this list, but they're not themselves in scope.
• Frameworks that don't support tool use or MCP directly. A wrapper that only does chat completion isn't an agent harness.
• Generic LLM client apps (chat UIs, prompt playgrounds) without tool execution.
• Agent-builder SDKs that only target cloud deployment (LangGraph Cloud, hosted CrewAI deployments, etc.) — the SDK itself may be fine if it's commonly run on desktop, but cloud-first framings are out.

Grey areas

These get judged case by case. If you're adding one, flag it in the PR description:

• IDE extensions with a usable CLI mode. Included if the CLI is first-class, excluded if it's a token-passing companion.
• Computer-use agents (Agent-S, UI-TARS, openfang). Included — they're the most "desktop-native" category there is — but they get their own section.
• Multi-agent frameworks that can run either locally or in the cloud. Included if the local mode is real and documented.
• Browser automation tools. Included only if they're explicitly designed to be driven by an on-device agent (not general Playwright/Puppeteer wrappers).
• Vendor CLIs that are thin clients to a hosted backend. Included if they do meaningful local tool execution; excluded if they're just a shell over a chat endpoint.

What the badges mean

See the badge legend in the README. The badges are descriptive, not gatekeeping — an entry can be in scope without having every badge filled in. When in doubt, leave a badge off rather than guess.

Updating this document

If you're unsure whether something belongs and this file doesn't answer the question, open an issue or PR to extend the grey-areas section. The goal is that the list stays coherent over time, not that it stays frozen.