Robot Skill Hub 0.5.2 — reusable robotics code by task

Reusable robotics code by task

Find a pinned implementation, inspect its I/O and source-specific findings, and prepare it in your own environment.

• 98 preparation recipes across eight runtime families, with current execution scope shown in the coverage report.
• Three small agent entrypoints for motion planning, kinematics and trajectory generation, delivered as checksum-verified HTTPS archives of about 23 kB each.
• Public HTTP search and three read-only MCP tools; discovery requires no Hub account or simulator.
• Exact-source findings and scoped experience sharing; canonical source, community reports and execution evidence remain distinct.
• Daily rolling access observations with thirty-day aggregate artifacts.

npx skills add https://hub.superfastrobotics.com --skill robot-motion-planning

Use robot-kinematics or robot-trajectory for the other task entrypoints. Installation fetches the selected small package without cloning Git. Preparing and running robot code remain separate local steps.

Agent contract · HTTP skill index · Task examples · Problem reproductions · MCP endpoint

Discovery and validation

Agent Skill Hub now carries all four packages, including the full CLI. In a fixed four-query sample on that provider, the Hub appeared in none of the results before the task-description update and in the first result for all four afterward: robot motion planning, inverse kinematics, trajectory generation, path smoothing. This is a small provider-search observation, not autonomous agent selection, general search accuracy or user adoption. The before/after receipt is attached.

The actual skills CLI installed all three packages over the public Hub HTTPS endpoint. Archive digests and every installed file matched the source; installed clients performed real search and pinned inspection. Maintainer tests opted out of access and installer telemetry. CI and deployment checks passed. Full ROS/RAI sessions and physical robot execution are not established by these checks.

Receipts are attached for the public HTTP installation, provider manifests and fixed search sample. Upstream contribution review and independent downstream reuse remain separate outcomes.

v0.4.3 — Reuse the fix. Keep the context lean.

Reuse the fix. Keep the context lean.

CLI inspection JSON: 17% fewer tokens. Same fields. All 17 recipe responses retained identical information; text tokens fell from 66,078 to 54,463 with o200k_base. The headline rounds 17.58% down. This measures CLI response text, not whole-task cost.

Two-link IK regression: 12/24 → 24/24 targets passed after saved-fix reuse. A controlled authored CPU fixture exercised actual local Worker/D1 storage, anonymous exact-scope recall, a separate reader process and linked outcomes. It does not establish autonomous LLM improvement or real-robot performance.

Measurements, every case and reproduction

Security fixes validate complete local/remote community content before delivery or hashing; malformed receipts preserve local data; invalid environment filters and coerced enrollment/parent types are rejected; patch bounds count UTF-8 bytes. Local validation passed 576 checks, with four Windows symlink-privilege skips. Linux CI and Pages/index activation passed. Hosted readback confirmed Worker 0.4.3, source searches, MCP inspection, seven invalid-filter rejections and existing finding recall.

Try account-free search · Agent entry · Read-only MCP

v0.4.2 - ASPIRE, real-platform sources and community robustness

Add 32 frozen source references from NVIDIA ASPIRE (13), Franka libfranka (9) and Hello Robot Stretch Body (10), including real-platform examples and agent improvement guides. The public catalog now contains 17 preparation recipes and 1,330 source references.

Community writes now enforce storage ceilings with explicit transactional guards. CLI recall preserves usable local findings when other records are malformed, and community patches reject agent-instruction targets. Operators upgrading their own Worker must apply migration 0004 before deployment.

Validation: 355 Python passes (4 Windows symlink skips), 102 API unit passes, 53 native workerd/D1 passes and 60 publication/sync passes. A controlled two-link IK fixture improved from 12/24 to 24/24 held-out targets after retrieving a known correction through the local community API. This verifies transfer and outcome linkage; it does not establish autonomous LLM improvement, token savings or robot execution.

• Catalog and agent entry
• Public read-only MCP
• Source pins and setup limits
• Local verification receipt

New upstream references retain source_reviewed/not_run. No new simulator or physical robot execution was performed. No paid plan was selected.

Robot Skill Hub 0.3.4

Agent discovery and contribution improvements for Robot Skill Hub.

• Search handles A*, D* and RRT* consistently across the CLI, HTTP API and read-only MCP tools.
• Catalog: 17 executable recipe entries, 1,262 source references and three reviewed findings. Seven additional PythonRobotics algorithm sources include immutable revisions, checksums and license provenance.
• Default offline bundles contain the published catalog generation 9cc6d2900d36d695a74a98a81ede20a77f60be54; bundled provenance intentionally identifies that generation rather than the later packaging commit.
• Source acquisition preserves manifests and supports interrupted retrieval recovery. Contribution errors exclude the active credential; offline submit/status commands stop before authentication or network access.
• Catalog activation reuses verified page memberships across publications, reducing repeated D1 writes. Free hosting remains in use.

Validation: Linux GitHub Actions structure/contract checks, Python tests, API unit/native-worker tests and live HTTP/MCP discovery readbacks passed for the published source. Local Windows Python validation passed 256 tests with four existing privilege-dependent skips.

Source references are metadata and acquisition instructions, not bundled simulator executions. The seven added upstream algorithms have not been executed by this project. Physical robot compatibility is not established by catalog inclusion. See each record's runtime and environment evidence.

Catalog: https://hub.superfastrobotics.com/
MCP: https://api.superfastrobotics.com/mcp

Robot Skill Hub 0.3.3 public preview

Robot discovery now includes 12 pinned MAVSDK and 18 Spot SDK Python entrypoints. The 1,206 existing official source records are preserved exactly. All 30 additions remain source references without a Hub execution adapter or inferred runtime compatibility; Spot's nonstandard upstream terms stay explicit.

An already-authorized finding save without a token now retains the finding locally and returns private credential setup guidance plus an explicit one-finding flush command. Duplicate saves remain read-only, and current project permission is checked before recovery guidance is shown.

Validation includes 190 Python tests (two existing Windows privilege skips), 34 publication checks, five wire checks, 11 sync checks, TypeScript checking, 87 Worker unit tests, native Workers tests and a successful native mirror of the actual publisher output. A fresh anonymous clone passed 240 checks and 19 CLI cases, with complete 1,254-document and three-finding API/static parity. All 128 verification requests opted out of access metrics. Three synthetic local save recovery cases made no network request and executed no flush.

This preview pins cbe459404040a145b32f2812b8b72ad93e97439c. Bundled offline snapshots honestly retain source commit e138db5a829f20d93165cc72f78c76548075afbe.

• Source review
• Checks
• Pages and index synchronization
• Hub and public API

No new simulator, model or physical robot execution was performed in this release. The launch access target is measured separately; maintainer verification opts out of that counter.

Robot Skill Hub 0.3.2 (public preview)

Agent discovery now finds declared robot types and model names consistently across the indexed API, static fallback and offline CLI. The catalog adds 47 pinned workflow source references from Isaac GR00T, RAI, OpenVLA, Diffusion Policy, ACT and Octo.

• Inventory: 17 preparation recipes, 1,207 source references and 3 reviewed findings. Source workflow coverage is 103 entrypoints across 11 upstream projects; it is a finite selection.
• Repairs: humanoid locomotion, G1, Crazyflie and OpenPI inference lookup; checksum-backed robot-name declarations; guarded literal console-entrypoint validation; fresh-cache retrieval of separately pinned discovery evidence.
• Requirements: preserve separate Python environments, code/model/dataset terms and 12 unresolved license states. All 47 additions remain source-only with no Hub preparation adapter and runtime_verification: not_run.
• Verification: 181 Python tests (2 Windows privilege-dependent cases skipped locally), 34 publication tests, 5 wire-format tests, 11 synchronization tests, TypeScript checks, 87 Worker unit tests and native workerd checks. The complete current publisher fixture passed all 14 catalog integration cases.
• Production: Checks, Pages and index synchronization, and official MCP Registry publication succeeded. The indexed generation contains all 1,224 documents and 3 findings. A fresh anonymous public checkout passed 230 checks and 16 CLI invocations; all verification requests opted out of campaign metrics.

The offline snapshots pin their actual 3a25e251c3b599413b7de3039a87618ba3cee72f generation. Existing recipe execution evidence remains record-specific. This release adds no new simulator, physical robot or model execution, and changes no paid plan.

Catalog · Public API · Read-only MCP

Robot Skill Hub 0.3.1 public preview

Preserves optional Python and CPU architecture within submitted environment combinations, and accepts exact positive JSON-safe artifact-size metadata without fetching the artifact. Intake requests remain limited to 16 KiB and publication still requires an internal draft PR and review.

Adds 99 pinned source references: 56 official training/inference/hardware workflow entrypoints and 43 GaP instruction documents. The inventory is 17 preparation recipes, 1,160 source references and 3 reviewed findings. Every new source remains not_run with no Hub preparation adapter; 11 conflicting upstream license declarations remain NOASSERTION/unverified. This is a finite source census.

Publishes complete provenance-manifest links alongside bounded search/findings discovery. Offline snapshots pin the actual 4608621 source commit and remain distinct from the online generation.

Validation: Linux Checks and Pages/index synchronization at 88022cc; 27 public HTTP checks; official MCP SDK 1.32.1 connection and three read-only tool calls; native Workers/D1 interoperability with the actual publisher output. No new robot, model, checkpoint or physical-hardware execution is claimed. Maintainer checks opt out of launch metrics.

v0.3.0: indexed discovery and robot artifact sources

Agents can search the complete published inventory and read selected sources without downloading whole catalog files.

• Native D1 FTS5 and compatibility indexes; bounded, checksum-bound static pages and atomic search/finding generation activation.
• Exact source/version/environment finding queries, per-receipt publication proofs, and CLI reads that preserve existing project locks and honest offline coverage.
• 17 preparation recipes, 1,061 pinned source references and 3 reviewed findings. The source references now include 409 robot description, model, scene and controller/configuration text artifacts from official repositories.
• New artifacts received static source/license inspection only. No new simulator, model-loading, Xacro evaluation, YAML semantic or physical-robot verification is claimed.

Validated locally with Python/catalog contracts, actual native D1/FTS publisher interoperability, and live public HTTP plus the official MCP SDK. GitHub Checks and Pages publication succeeded for the pinned source commit. MCP remains read-only; local preparation uses the separate Hub CLI. Public preview, on the existing free plans.

Hub: https://hub.superfastrobotics.com/
API: https://api.superfastrobotics.com/v1/search
MCP: https://api.superfastrobotics.com/mcp

Robot Skill Hub 0.2.1: precise filters and CLI preparation clarity

Robot and runtime filters now behave consistently across the installed CLI and public API. Reference-only discoveries honor their declared robot/framework metadata; multiple environment filters still require one documented compatibility combination. Multiword queries require every term, and unfamiliar Unicode text no longer silently becomes an empty query.

Preparing a source reference now returns source_reference_no_adapter with upstream source/setup links and creates no project environment. The CLI also defaults authenticated contributions to the canonical Hub API endpoint.

The MCP Registry description now explicitly states that local preparation uses the separate Hub CLI. This corrects ambiguity identified by the Commons host Hermes after it independently checked the deployed MCP and JSON contracts. Original feedback.

• Active MCP Registry 0.2.1
• Checks passed
• Pages publication passed
• Registry publication and public readback passed

Source commit: 1297f636a24da0daeec9432d42d7449030d1ac0b. The refreshed offline snapshot contains 17 preparation recipes, 652 pinned source references and 3 reviewed findings. The reviewed OpenPI policy-client submission adds one discovery reference; it does not add a Hub runtime adapter.

The three MCP tools remain read-only. Robot task evidence remains limited to the two initial Panda recipes at seed 0 under the explicit Hub 1000-step budget; this patch adds no new robot execution claim. Agent contract ? Quickstart.

Robot Skill Hub 0.2.0: public JSON and MCP discovery

Discover pinned robotics code through public JSON search or read-only MCP, inspect its exact sources and evidence, then prepare supported recipes locally.

• Public JSON search: https://api.superfastrobotics.com/v1/search?q=pick%20cube
• Read-only Streamable HTTP MCP: https://api.superfastrobotics.com/mcp (find_robot_skills, inspect_robot_skill, get_hub_capabilities). Discovery needs no account.
• Agent contract and instructions: https://hub.superfastrobotics.com/agent.json
• Reviewed task findings can be reused across exact recipe/source versions. Public saves require an authorized project scope and scoped token, and enter an internal draft PR for maintainer review.

This publication contains 17 preparation recipes, 651 pinned source references and 3 reviewed findings. Source references retain their upstream license, provenance and invocation context; they do not have Hub runtime adapters. Prepared recipes and source references have separate evidence levels.

Actual simulation evidence is scoped to Panda PickCube and StackCube, seed 0, CPU/state execution, with an explicit 1000-step motion-planning budget and checked truncation flags. The remaining recipes and source references are not claims of verified task execution. This does not establish the upstream 50-step benchmark result or physical hardware operation. Exact execution evidence.

Install the agent skill into a project:

npx skills add Superfast-Robotics/robot-skill-hub --skill find-robot-skills -a codex -y

Use -a claude-code for Claude Code. Quickstart ? Catalog ? Contribution contract.

Public MCP Registry record: https://registry.modelcontextprotocol.io/v0.1/servers/io.github.Superfast-Robotics%2Frobot-skill-hub/versions/0.2.0

Release source: 8adbd359627f39c68b49b74bbb8b97fac617573b. This is a public preview; compatibility and verification remain record-specific.