Live project registry for coding agents — projects, skills & prompts over MCP. Exposes local project/skill/prompt registry to coding agents via MCP, with read tools for context and write tools for keeping the registry alive.
KeyStack has 12 well-named tools with clear verb-noun patterns (list_, get_, create_, update_, search_, upsert_). Descriptions are present and contextual (avg ~150 chars), explaining WHEN to use each tool. Input schemas are visible and mostly complete with type declarations. However, many parameter descriptions are minimal or missing entirely (e.g., 'slug' in get_project has no description; 'path' in scan_repo lacks format guidance). Output schemas are not documented, LLMs cannot predict what fields to expect from responses. Error handling is absent from tool definitions. The server handles both read-only and write operations but lacks confirmation patterns for destructive tools (create_project, update_project, upsert_skill, upsert_prompt). No tool annotations (readOnlyHint, destructiveHint) are present.
Add a new project to the registry.
Full context for one project by slug: description, stage, stack (language/frameworks/db), services, tests, next steps, github, keys_ref, dev-state (track/health_score/has_*), and its .ai-codex spec registry (specs[], open_specs). Use at the start of work on a project.
Get one prompt by slug, including its full body.
Get one skill by slug.
Overview of all projects in the registry (name, slug, stage, stack, services, last touched, track, health_score, open_specs). Call this to see the whole portfolio map.
List all stored prompts (name, description, category).
Output schemas not documented. LLMs cannot predict response structure (fields, types, pagination). Forces agents to guess what data is available after each call.
Parameter descriptions missing or minimal. 'slug' appears in 6 tools with no description; 'path' in scan_repo lacks format guidance (absolute vs relative?); 'query' in search_projects has no examples or constraints.
No error handling guidance in tool definitions. No recovery hints (e.g., 'If project not found, try search_projects'). No categorization of errors as retryable vs fatal.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 64 | 2026-07-28+ | v2 |
List all stored skills (name, description, what it does, location).
Inspect a local project folder and propose registry fields (language, frameworks, database, services, github_url, tests). Read-only — does NOT save. Use it, then confirm and call create_project / update_project with the results.
Find projects by free text across name, language, frameworks, database and services. E.g. 'which project uses Resend?'
Update fields of an existing project — only pass what changed. Use this to keep the registry current: stage, tests, next steps, stack, services.
Add or update a reusable prompt in the registry.
Add or update a skill in the registry (so you remember what skills you have).
Destructive tools (create_project, update_project, upsert_skill, upsert_prompt) lack confirmation or dry-run patterns. No tool annotations (destructiveHint) to signal irreversibility to LLMs.
Parameter constraints not formalized. 'stage' enum is declared in create_project/update_project, but 'type' accepts free-form strings ('mobile | web | saas | bot | cli | api | library | desktop') without enum constraint. Invites hallucinated values.