A resident, cross-tool skill brain for AI coding agents. Routes every task through one shared memory: reuse local skills, recall your notes, research curated sources, or build a new skill. Works across Claude Code, Codex, and more over MCP.
Mycelium defines 5 tools with explicit schemas and descriptions. All tools have non-empty descriptions (range 150 - 450 chars, within baseline 34 - 392). All parameters are typed (Zod schemas visible in server.ts). However, output schemas are NOT documented, responses are JSON strings without field-level documentation. Error handling is minimal (validation errors only, no recovery guidance). Tool naming is domain-specific but not verb-first (consult, feedback, register_skill follow patterns; catalog_ingest/catalog_assess are compound nouns). Parameter descriptions are present but sparse, many lack format/constraint details. No tool annotations (readOnlyHint, destructiveHint, idempotentHint). Overall: solid foundation, but lacks LLM-optimized descriptions, output documentation, and error recovery patterns.
Submit a batch of SEMANTIC SECURITY VERDICTS for already-cataloged skills (a body-reading audit pass). For each skill you READ the full SKILL.md and report EVIDENCE — never a safety decision; Mycelium decides the tier. `klass`: "performs" (the body actually runs/instructs a dangerous capability), "detects" (it audits/scans/teaches detection OF danger — not itself dangerous), or "discusses" (merely mentions danger in prose). `caps`: which capability labels the body actually invokes, from this controlled set — pipe-to-shell, reverse-shell, rm-rf-dangerous, privilege-escalation, credential-exfil, financial-action, exfiltration, prompt-injection (severe); dynamic-exec, installs-packages, network, credential-handling, base64-decode (soft). `evidence`: a short quote. Match name+purpose EXACTLY as ingested (hash key). Returns {applied, notFound, byTier}.
Ingest a batch of ecosystem skills into the knowledge catalog (NOT install them). For a crawl of anthropics/skills, antigravity-awesome-skills, skills.sh, or skillsmp. Mycelium risk-classifies and tiers each entry on its side — you supply only raw metadata (name, purpose, url, keywords, stars, optional scanText). De-dupes by content. Returns {inserted, duplicates, byTier}.
MUST be called before starting any task. Returns a verdict: reuse (a local skill matched), recall (your memory covers it — read the notes), install (no local skill, but the catalog knows ready-made ones — SHOW the user each candidate by name + source + url + one-line purpose and ASK if they want it installed; do not silently swallow it or research the web instead), searching (catalog empty too, researching online), build (recurring need with no coverage — interactively build a skill then call register_skill), or pass.
Output schemas undocumented. Tools return JSON strings (consult, catalog_ingest, catalog_assess) without field-level documentation. LLMs cannot plan downstream calls or extract structured data reliably.
Tool names lack action verbs. 'consult', 'feedback', 'register_skill', 'catalog_ingest', 'catalog_assess' are domain-specific but not verb_noun pattern (get_, create_, update_, search_). LLMs infer intent slower from non-standard names.
Error handling lacks recovery guidance. Validation errors return raw Zod messages ('invalid consult: ...') without actionable next steps. No categorization (retryable vs user-fixable vs fatal).
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 56 | 2026-07-28+ | v2 |
Report a skill outcome so the brain learns. Pass `task` (what the skill was for). Outcomes: "ok" (worked — strengthens it and clears any misfit for this task-shape), "fail" (used but failed — weakens it and records a misfit), "reject" (the suggested skill was IRRELEVANT to the task — does NOT weaken the skill, only records a skill×task-shape misfit so it stops being suggested for THIS kind of task). Call "reject" the moment a reuse suggestion is off-topic.
Deposit a skill you built interactively (after a "build" verdict): pass the full SKILL.md and a one-line purpose. Also pass `keywords`: the trigger task's key terms IN THE USER'S LANGUAGE (e.g. Chinese 中断/固件/MIDI) so the SAME task that prompted the build can find this skill later — without them an English-only skill is unmatchable by a Chinese task. The brain lands it, annotates its purpose so other tools know what it is for, and de-dupes by name.
Parameter descriptions lack format/constraint details. 'task' is 'short description' (vague); 'skillMd' is 'full SKILL.md content' (no length/format spec); 'keywords' is 'trigger task key terms' (no language guidance). LLMs cannot validate inputs.
No tool annotations. Tools lack readOnlyHint (consult is read-only), destructiveHint (feedback, register_skill, catalog_ingest, catalog_assess modify state), or idempotentHint. Agents cannot reason about side effects.