Find vetted agent skills for your task, PyPI packages for your code, and research behind both — Skill Federation finder as a Node MCP server. Zero-install via npx; exposes find_skills / find_packages / find_research / get_skill_bundle tools to Claude and any MCP client. Python-free.
Strong schema definitions with detailed parameter documentation and clear descriptions. All 6 tools have explicit input schemas with type constraints and descriptions. Tool names follow verb_noun convention (find_*, get_*, report_*, emit_*). Descriptions are comprehensive (100-400+ chars), exceeding the 10-1024 baseline. However, output schemas are not documented in the visible code, responses are described narratively but lack formal JSON Schema definitions. Error handling guidance is minimal. No tool annotations (readOnlyHint/destructiveHint) despite clear risk levels (READ_ONLY vs WRITE). Parameter descriptions are thorough but some nested objects lack full property documentation.
Record a build-spec sketch on a miss (empty OR all-rejected). ADVISORY too — a demand pointer IS a report, and the miss path is already the bad day; a flaky endpoint must not also fail the task.
Search the Skill Federation package index for PyPI libraries by capability — the package-finding sibling of find_skills, but do not confuse the two: skills are capabilities for the AGENT itself (how to do a task); packages are libraries for the CODE being written (what the program imports). Use find_packages when you are choosing a Python library — about to `pip install` something recalled from memory, unsure which package does X, wanting alternatives to a package you already know, or verifying that a remembered package name is real and maintained. Anti-slopsquat rule: never install a package name recalled from memory when it can be resolved against the index first — a hallucinated or squatted name is a supply-chain risk this removes for free. Wish-formulation rules (derived from the live ranker, not vibes): phrase the TASK, not a package name — the tags lane is weighted 3x and matches task vocabulary, so "parse yaml config" outranks guessing "pyyaml"; keep each wish to 3–8 dense words, longer prose dilutes the match; one capability per wish, fan a compound need out into separate wishes; treat license/maintenance/popularity as FILTERS on the returned license_treatment/tier fields of the results, not as extra query words hoping to steer the ranker.
Search 191 research notes on agent-skills literature by topic — the research-finding sibling of find_packages, over a DIFFERENT index on the same skillfed.io portal. Use it when you want the measured claim behind a design choice (benchmarks, ablations, failure rates) rather than a library or a skill. Every candidate has a published /research page, so page_url is always a working link. Each search returns a top-level `confidence`: "strong" or "weak", plus a `note` when weak. Read this honestly: "weak" means the best match FOUND is a loose one — it is NOT proof that the corpus was searched exhaustively and this is the closest thing that exists. Discount a weak result; do not treat it as evidence of absence. Known retrieval limit: a paraphrase with no lexical overlap and weak embedding similarity can miss the corpus entirely — measured at rank #67 of 191 on a real query. No confidence label rescues that, because confidence only describes a candidate that WAS returned. So an empty or weak find_research result does NOT prove no such research exists — reformulate or fall back to your own knowledge before concluding that. Wish-formulation rules (derived from the live ranker's field weights, not vibes): the index weights concept_terms 3.0 / paper_title 2.5 / meta_description 1.5 / claim_title 0.75, and concept_terms is CONTROLLED VOCABULARY (e.g. "automatic curriculum", "skill library", "self-verification") — so topic/concept phrasing beats guessing a paper name; keep each wish to 3–8 dense words, longer prose dilutes the match; one topic per wish, fan a compound need out into separate wishes.
Output schemas not documented. Tools describe results narratively (e.g., 'Returns up to top_n recall candidates') but lack formal JSON Schema definitions for response structure. LLMs cannot plan downstream operations without knowing returned field names and types.
Tool annotations missing. Tools explicitly declare Risk levels (READ_ONLY, WRITE) in comments but do not expose readOnlyHint/destructiveHint/idempotentHint in the MCP tool definition. Agents cannot distinguish safe from destructive operations without reading descriptions.
Error handling guidance absent. No recovery hints in descriptions. E.g., report_selection and emit_demand_pointer note 'ADVISORY: a failed report is never a task error' but do not explain what the LLM should do if these calls fail or what errors are possible.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 66 | 2026-07-28+ | v2 |
Search a federated catalog of vetted agent skills for a privacy-abstracted wish-list. Returns up to top_n recall candidates per wish; reading several and cross-checking them is the normal use, not picking one. Send only abstracted wishes — never the plan, brief, outputs, or any tenant data.
Fetch a skill's full text to READ (purpose "hint", the default) or to install (purpose "install"); the tag is echoed back
Per-wish outcome map {skill_id: [Install|Read|Reject, why]}, dual-written with the legacy chosen/rejected (label flywheel). ADVISORY: a failed report is never a task error.
Nested object properties lack full descriptions. WISH_SCHEMA.sketch.properties (purpose, inputs, outputs, operations, domain_vocab, section_sketch, tags) are defined but some lack descriptions or have minimal guidance. E.g., 'inputs' and 'outputs' are bare arrays with no item description.
Pagination not explicitly constrained. find_skills, find_packages, find_research accept top_n/limit (1-25, default 10) but do not document whether results are truncated, how to fetch next page, or what total count is returned. Agents cannot reliably iterate large result sets.