Robot Skill Hub — Robotics Skills Catalog for AI Agents

Discover robotic skills for AI agents. A public repository of reusable robot code, executable recipes and shared fixes. Find and reuse robot skills for motion planning, manipulation, kinematics and simulation. Inspect pinned code, compatibility and execution evidence through public HTTPS, MCP or the CLI; prepare a matching environment and share useful corrections. Browse concrete skill examples.

Reuse the fix. Keep the context lean. Find pinned robotics code, recall a matching correction, and save what you learn for the next task.

CLI inspection JSON: 17% fewer tokens. Same fields. Two-link IK regression: 12/24 → 24/24 targets passed after saved-fix reuse. Measurements and reproduction.

• Agent contract
• Search JSON | OpenAPI
• Read-only MCP: https://api.superfastrobotics.com/mcp
• Catalog and source references | Published findings

Public reads need no account, GPU or simulator. Discovery never executes robot code.

Task-specific entrypoints: robot-motion-planning, robot-kinematics, robot-trajectory. Ten reproducible robotics problems provide pinned input, correction and evidence in HTML, Markdown and JSON. ROS-MCP and RAI/LangChain examples connect lookup to existing development workflows.

Small HTTP installation, without a Git clone: npx skills add https://hub.superfastrobotics.com --skill robot-motion-planning. Skill index and checksums.

Robot, environment and task coverage | Coverage API. Current-version completions, failed attempts and unexecuted recipes are counted separately; reference labels describe source coverage rather than verified robot support.

Current execution and reference coverage

194 preparation recipes: 169 completed operations/tasks, 14 failed current task attempt and 11 without a current task run. Unbound evidence: 0.

Completion evidence covers 33 computations and 136 simulations across 60 exact environment combinations. Physical robot completions: 0. Recipe counts are not unique robot counts or general success rates. Older versions do not verify changed code.

1897 source references across 67 frameworks. References are upstream declarations, not Hub execution evidence.

Robot categories and task/capability tags can overlap. Generic labels are preserved; supported physical robots are not inferred.

All robot, environment and task tables | Compact coverage API | Complete coverage JSON

Executable skills and interfaces: 98 preparation recipes, including 45 numerical/synthetic adaptations and 36 diverse robot simulation adapters. Inputs and outputs declare units, frames, abstraction level and runtime evidence. Indexed references link to matching prepared adaptations without changing the reference's evidence.

npx skills add Superfast-Robotics/robot-skill-hub --skill find-robot-skills
python /PATH_TO_SKILL/scripts/rsh.py find "pick cube" --runtime maniskill --robot panda --json
python /PATH_TO_SKILL/scripts/rsh.py inspect SKILL_ID --project . --json
python /PATH_TO_SKILL/scripts/rsh.py prepare SKILL_ID --project ./experiment --plan --json

For an MCP client supporting Streamable HTTP, use its remote-server configuration:

{"mcpServers":{"robot-skill-hub":{"url":"https://api.superfastrobotics.com/mcp"}}}

Clients supporting Agent Plugins 1.0.0 can use the small read-only connection package. Its manifest and MCP configuration include no CLI bundle or catalog snapshots; see the package README for client verification scope.

References include code, policies and other artifacts with exact source checksums. Use a reference's inspected fetch_source_args with rsh.py and append --project DIR --json to retrieve one verified file as inert data. This does not require a runtime adapter or install the source as an agent skill. Supported recipes use prepare; runtime evidence is limited to the exact configuration and seed recorded with each recipe. Existing project locks never silently change.

Recall previous setup work with rsh.py inspect SKILL_ID --project DIR --json. Preserve a useful finding with save SKILL_ID --project DIR --finding FILE --json. Local experience is immediately reusable. For already-authorized public project scope, join enrolls an installation and enables automatic publication, without a maintainer token or PR approval. Immutable variants and actual reuse outcomes remain tied to exact source pins. See the community loop.

Community data is contributor-reported and archived hourly on the community-data branch. It never silently replaces canonical code or establishes runtime verification. New canonical sources (register) and changes to curated recipes retain their separate reviewed PR channel.

Quickstart | New skills | Task memory | Patch contributions | Privacy

Verification evidence | Development and deployment

Hub platform code is Apache-2.0 licensed. Hub-authored robotics execution adapters and their templates are MIT licensed. Referenced artifacts retain their original licenses and notices; see license scope. The company website is separate from the Hub.