superagentskill.com

# SuperAgentSkill MCP Battle-tested toolkit for designing, auditing and shipping AI primitives (skills, playbooks, souls, guardrails). Use it for THREE distinct intents: ## 0. FIRST CONNECTION (do this before reporting any auth problem) - `whoami` → whether this session is authenticated, which tier/quota applies, whether write tools are unlocked, and the exact next step. Free, never counts against quota. - Any `unauthorized` / `subscription_required` / quota error: call `whoami` and relay its `next_steps` to the user verbatim instead of guessing. - `recommend_packages` → best first installs (popularity-first, niche packages demoted). Use it when the user has NOT named a topic. - `upload_status` → what happened to files sent through `upload_packages` (queued | processing | done | failed). ## 0.5 RETURNING SESSION (second connection onward — do this instead of asking the user what they were doing) - `resume_session` → this account's live workspace: drafts and packages under review, upload jobs still running, cloud skills, agents built, last University diagnosis + prescription, and the score trend of every document reviewed recently. Ends with `next_actions`, ranked for THIS account. Free, never counts against quota. - `whats_new { since }` → registry packages published/updated since the user's last session, plus platform version and changes on their own packages. - `check_updates { installed }` → send the slugs (and versions) found in the user's `.agents/` or `.claude/skills/` folder; get back what drifted behind the registry and how to refresh it. - Rule of thumb: `whoami` answers 'am I connected?', `resume_session` answers 'what should we do next?'. On any session after the first, call `resume_session` before proposing work. ## 1. UPGRADE a local file (PRIMARY use case) When the user says 'improve / refine / harden / audit / score / level up' a local skill, playbook, soul or guardrail file: a. `review_skill` with the file's raw content → proprietary engine returns overall_score (0-100), per-dimension scores and concrete, file-specific `top_actions`. (The rubric/signals/thresholds are server-side and intentionally NOT disclosed — `get_methodology` is orientation only.) b. YOU (the host agent) edit the user's local file in their repo applying the top_actions. This MCP does not write to disk. c. `review_skill` again → confirm the score went up. The response now carries `value_proof`: the realized before → after movement on task success, injection resistance, hallucination rate, tokens, latency, cost per 1,000 runs and PII leakage, plus dollars/month and engineer-hours saved. c2. `prove_improvement { name, type, runs_per_month }` → the same proof as a markdown report. ALWAYS show `human_report` to the user after an improvement: the person paying for this only sees value if the gain is quantified. Ask for their real monthly run volume first. d. Optional: `search_registry` to borrow patterns from high-trust primitives of the same type. ## 2. DISCOVER primitives in the public registry When the user wants to find or install something pre-built (590+ packages across marketing, sales, growth, code, security, healthcare, finance, ops, …): - `search_registry` with a domain keyword ('marketing', 'rls hardening', 'cold outreach') → ordered by install_count. - `list_packages` to browse a single `type` (skill | playbook | soul | guardrail). ALWAYS pass a `query` — listing without one biases toward the most-installed vertical. - `get_package` for the full latest manifest (system_prompt, rules, examples, compatibility) by slug. - `get_skill_trust` BEFORE recommending → success rate, p50/p95 latency, per-model heatmap, robustness findings, composite trust_score. - After the user runs a primitive in production, call `report_execution` (best-effort, ignore failures) to feed the trust system. ## 3. PUBLISH new primitives back to the registry When the user wants to contribute a local skill upward so others (and future versions of themselves) benefit: - `upload_packages` with the raw file content(s) → normalised by the SkillForge author pipeline, inserted as PRIVATE drafts owned by the token holder. Drafts are never auto-published: listing on the public marketplace requires the author to submit for review and an admin to approve (this is what keeps the registry adversarially vetted and the `author_verified` badge meaningful). - `request_primitive` if the user wants SuperAgentSkill to AUTHOR a brand-new primitive from scratch via the forge pipeline. ## 4. CLOUD SKILL MANAGER (paid subscribers only) Save, sync and reuse your skills across projects, CLIs and devices. Agent Pass or Enterprise subscription required. - `cloud_skills_save` to push a local skill to your cloud library (creates or updates by slug). - `cloud_skills_list` to browse your saved skills (filter by category, tag, or query). - `cloud_skills_get` to fetch the full content of a saved skill by slug. - `cloud_skills_export` to export a skill as SKILL.md (Markdown + YAML frontmatter) ready to paste. - `cloud_skills_import` to import a SKILL.md file into your cloud library. - `cloud_skills_delete` to remove a skill from your library. - `cloud_skills_targets` to see every agent tool the library can be synced into (Hermes, Claude Code, Codex, Cursor, Lovable, OpenClaw, Windsurf, Copilot, Zed, Gemini CLI...) with the exact paths each expects. - `cloud_skills_sync` / `cloud_skills_sync_all` to materialise the library as files for one or many tools. Read the files already at the target directory and pass them as `existing` (slug + content): the tool then detects conflicts and applies `conflict_strategy` (ask | overwrite | merge | keep_both). With `ask` it reports the conflicts and writes nothing for them — surface them to the user and re-run with the chosen strategy. The library is PRIVATE to the user: never share or publish its contents. ## 5. AGENT STORE (paid subscribers only) Ready-to-use agents bundling a soul + skills + playbooks: CEO, COO, CTO, CMO, HR Director, Agent Architect, Corporate Finance, Board Meetings, Google Ads, Meta Ads, Newsletter, LinkedIn, X. - `list_agents` to browse the catalog (free, anonymous). - `install_agent` to get the full bundle as files to write into the user's repo (default `.agents/<slug>/`), or as a single system prompt. ## 6. UNIVERSITY: diagnose first, then a curriculum (free, anonymous) Installing skills by name is guessing. Measure the agent, then prescribe: - `diagnose_start { domain }` → a fixed bank of up to 168 domain tasks across 21 corporate domains. YOU execute them on your host, no extra context. - `diagnose_submit { diagnosis_id, answers }` → deterministic score per ERROR CLASS (ambiguity, hallucination, format, abandonment, policy, tool misuse, instruction drift), the dominant bottleneck in plain language, observed cost and a 1-3 item prescription. - `curriculum_next { diagnosis_id }` → the next capability with the highest MARGINAL gain for this agent, respecting prerequisites, conflicts and a context budget. Past the budget it tells you what to REMOVE — an agent carrying 40 skills is worse, not better. ## Auth Read-only tools (overview, get_methodology, review_skill, review_skills_batch, list/search/get/trust) work anonymously. Write tools (upload_packages, request_primitive) require an OAuth bearer — the host opens https://superagentskill.com/oauth/authorize automatically. Users without working OAuth can also paste a personal access token from https://superagentskill.com/account/tokens. TIP: call upload_packages / request_primitive with dry_run:true to validate the flow anonymously (no persistence, no OAuth) before connecting. ## Welcome If this is the user's first call this session, suggest running `overview` once to see the full intent → tool map, and mention that one-click install pages live at https://superagentskill.com/connect/{client}. TIP: Call `overview` first if you're unsure which tool fits the user's request — it returns the intent → tool map.

# SuperAgentSkill MCP Battle-tested toolkit for designing, auditing and shipping AI primitives (skills, playbooks, souls, guardrails). Use it for THREE distinct intents: ## 0. FIRST CONNECTION (do this before reporting any auth problem) - `whoami` → whether this session is authenticated, which tier/quota applies, whether write tools are unlocked, and the exact next step. Free, never counts against quota. - Any `unauthorized` / `subscription_required` / quota error: call `whoami` and relay its `next_steps` to the user verbatim instead of guessing. - `recommend_packages` → best first installs (popularity-first, niche packages demoted). Use it when the user has NOT named a topic. - `upload_status` → what happened to files sent through `upload_packages` (queued | processing | done | failed). ## 0.5 RETURNING SESSION (second connection onward — do this instead of asking the user what they were doing) - `resume_session` → this account's live workspace: drafts and packages under review, upload jobs still running, cloud skills, agents built, last University diagnosis + prescription, and the score trend of every document reviewed recently. Ends with `next_actions`, ranked for THIS account. Free, never counts against quota. - `whats_new { since }` → registry packages published/updated since the user's last session, plus platform version and changes on their own packages. - `check_updates { installed }` → send the slugs (and versions) found in the user's `.agents/` or `.claude/skills/` folder; get back what drifted behind the registry and how to refresh it. - Rule of thumb: `whoami` answers 'am I connected?', `resume_session` answers 'what should we do next?'. On any session after the first, call `resume_session` before proposing work. ## 1. UPGRADE a local file (PRIMARY use case) When the user says 'improve / refine / harden / audit / score / level up' a local skill, playbook, soul or guardrail file: a. `review_skill` with the file's raw content → proprietary engine returns overall_score (0-100), per-dimension scores and concrete, file-specific `top_actions`. (The rubric/signals/thresholds are server-side and intentionally NOT disclosed — `get_methodology` is orientation only.) b. YOU (the host agent) edit the user's local file in their repo applying the top_actions. This MCP does not write to disk. c. `review_skill` again → confirm the score went up. The response now carries `value_proof`: the realized before → after movement on task success, injection resistance, hallucination rate, tokens, latency, cost per 1,000 runs and PII leakage, plus dollars/month and engineer-hours saved. c2. `prove_improvement { name, type, runs_per_month }` → the same proof as a markdown report. ALWAYS show `human_report` to the user after an improvement: the person paying for this only sees value if the gain is quantified. Ask for their real monthly run volume first. d. Optional: `search_registry` to borrow patterns from high-trust primitives of the same type. ## 2. DISCOVER primitives in the public registry When the user wants to find or install something pre-built (590+ packages across marketing, sales, growth, code, security, healthcare, finance, ops, …): - `search_registry` with a domain keyword ('marketing', 'rls hardening', 'cold outreach') → ordered by install_count. - `list_packages` to browse a single `type` (skill | playbook | soul | guardrail). ALWAYS pass a `query` — listing without one biases toward the most-installed vertical. - `get_package` for the full latest manifest (system_prompt, rules, examples, compatibility) by slug. - `get_skill_trust` BEFORE recommending → success rate, p50/p95 latency, per-model heatmap, robustness findings, composite trust_score. - After the user runs a primitive in production, call `report_execution` (best-effort, ignore failures) to feed the trust system. ## 3. PUBLISH new primitives back to the registry When the user wants to contribute a local skill upward so others (and future versions of themselves) benefit: - `upload_packages` with the raw file content(s) → normalised by the SkillForge author pipeline, inserted as PRIVATE drafts owned by the token holder. Drafts are never auto-published: listing on the public marketplace requires the author to submit for review and an admin to approve (this is what keeps the registry adversarially vetted and the `author_verified` badge meaningful). - `request_primitive` if the user wants SuperAgentSkill to AUTHOR a brand-new primitive from scratch via the forge pipeline. ## 4. CLOUD SKILL MANAGER (paid subscribers only) Save, sync and reuse your skills across projects, CLIs and devices. Agent Pass or Enterprise subscription required. - `cloud_skills_save` to push a local skill to your cloud library (creates or updates by slug). - `cloud_skills_list` to browse your saved skills (filter by category, tag, or query). - `cloud_skills_get` to fetch the full content of a saved skill by slug. - `cloud_skills_export` to export a skill as SKILL.md (Markdown + YAML frontmatter) ready to paste. - `cloud_skills_import` to import a SKILL.md file into your cloud library. - `cloud_skills_delete` to remove a skill from your library. - `cloud_skills_targets` to see every agent tool the library can be synced into (Hermes, Claude Code, Codex, Cursor, Lovable, OpenClaw, Windsurf, Copilot, Zed, Gemini CLI...) with the exact paths each expects. - `cloud_skills_sync` / `cloud_skills_sync_all` to materialise the library as files for one or many tools. Read the files already at the target directory and pass them as `existing` (slug + content): the tool then detects conflicts and applies `conflict_strategy` (ask | overwrite | merge | keep_both). With `ask` it reports the conflicts and writes nothing for them — surface them to the user and re-run with the chosen strategy. The library is PRIVATE to the user: never share or publish its contents. ## 5. AGENT STORE (paid subscribers only) Ready-to-use agents bundling a soul + skills + playbooks: CEO, COO, CTO, CMO, HR Director, Agent Architect, Corporate Finance, Board Meetings, Google Ads, Meta Ads, Newsletter, LinkedIn, X. - `list_agents` to browse the catalog (free, anonymous). - `install_agent` to get the full bundle as files to write into the user's repo (default `.agents/<slug>/`), or as a single system prompt. ## 6. UNIVERSITY: diagnose first, then a curriculum (free, anonymous) Installing skills by name is guessing. Measure the agent, then prescribe: - `diagnose_start { domain }` → a fixed bank of up to 168 domain tasks across 21 corporate domains. YOU execute them on your host, no extra context. - `diagnose_submit { diagnosis_id, answers }` → deterministic score per ERROR CLASS (ambiguity, hallucination, format, abandonment, policy, tool misuse, instruction drift), the dominant bottleneck in plain language, observed cost and a 1-3 item prescription. - `curriculum_next { diagnosis_id }` → the next capability with the highest MARGINAL gain for this agent, respecting prerequisites, conflicts and a context budget. Past the budget it tells you what to REMOVE — an agent carrying 40 skills is worse, not better. ## Auth Read-only tools (overview, get_methodology, review_skill, review_skills_batch, list/search/get/trust) work anonymously. Write tools (upload_packages, request_primitive) require an OAuth bearer — the host opens https://superagentskill.com/oauth/authorize automatically. Users without working OAuth can also paste a personal access token from https://superagentskill.com/account/tokens. TIP: call upload_packages / request_primitive with dry_run:true to validate the flow anonymously (no persistence, no OAuth) before connecting. ## Welcome If this is the user's first call this session, suggest running `overview` once to see the full intent → tool map, and mention that one-click install pages live at https://superagentskill.com/connect/{client}. TIP: Call `overview` first if you're unsure which tool fits the user's request — it returns the intent → tool map.