Skip to content
UsecaseBoard
← All use cases
githubcomputer-use

OpenGCUA + MAI-Gym: Alibaba open stack for generalist computer use

At a glance

Tongyi-MAI open-sources OpenGCUA (GUI+code harness, harness-to-model decisions, cross-device SDK; OSWorld-v2 +19.6 pts / −53% steps across 8 models) and MAI-Gym (up to 20k parallel RL envs; MAI-Gym-9B 78.4% MobileWorld / 76.2% OSWorld-Verified). Code + arXiv:2610.22200 / 2610.20939. This workflow is listed under computer-use, open-source, GUI-agents, RL, OSWorld.

Agents / tools
OpenGCUA, MAI-Gym, Tongyi-MAI, computer-use agents

Building a tool for these workflows? Sponsor / inquire.

Save ideas in this browser. Upvotes count across the board.

Prompt excerpt

Clone Tongyi-MAI/OpenGCUA, wire the hybrid GUI+code harness for a desktop task, and compare step count vs a GUI-only baseline on the same OSWorld-style prompt.
Explore all →
docsMeta-Muse

metacua: Muse Spark operates a real Mac via native GUI computer-use

Meta’s official macOS cookbook ships metacua—a hackable CLI where Muse Spark captures the screen, calls computer tools with normalized 0–1000 coordinates, and executes native mouse/keyboard events until `computer.stop`. Distinct from community Muse Jailbreak (Chrome tunnel): this is Model API + local GUI loop. Warning: real Mac control; prefer a VM/dedicated user; leave `--allow-bash` off unless isolated.

Meta MuseMuse Sparkmetacua
Source ↗
githubobservability

Tracekit 1.2.3: tamper-evident flight recorder for AI agents

MIT Python package signs every policy-gated tool call into a hash-chained .tkb file verifiable offline (Python+TS); works across Claude Code, OpenAI Agents, LangChain/LangGraph, MCP and more—agents can edit their own logs but not hide edits when the signer is out of reach.

TracekitClaude CodeOpenAI Agents+3
Source ↗
githubmulti-agent

Claustrum: harness-neutral architect that delegates to any coding agent via MCP

Claustrum (alpha, .NET NativeAOT) is a harness-neutral conductor: an architect in Claude Code, Cursor, Copilot CLI, opencode, or desktop delegates builds/reviews/tests to agents on any other harness/model and gets one normalized JSON receipt. Two doors — `claustrum run` CLI and `claustrum mcp` stdio server. Parallel builders on git worktrees, budget ledger, backends doctor. Role-based routing (strong model for architect/reviewer, cheap for builders) without locking the pipeline to one vendor.

ClaustrumClaude Codeopencode+3
Source ↗
githubsandbox

DeltaBox Lite: diff/commit/abort sandbox for coding agents

Apache-2.0 agent-native sandbox (Linux FUSE/Landlock, macOS APFS clones+Seatbelt) gives agents a confined FS view with unified text diffs, binary markers, and per-task or per-tool-call commit/abort—optional JSON-RPC adapters for Pi and policy for secrets/network.

DeltaBox LitePicoding agents
Source ↗