Skip to content

knowledge-server.md

Latest commit

 

History

History
47 lines (32 loc) · 2.96 KB

knowledge-server.md

File metadata and controls

47 lines (32 loc) · 2.96 KB
Raw

Apple Knowledge MCP

Context

ios-agent-knowledge is the public-reference companion to ios-agent-mcp. Both binaries ship in ios-agent-mcp@2.4.0. The analyzer keeps its eleven local-project tools; the knowledge server has eight separate tools and never exposes project paths over HTTP.

Pattern

Tool Result
search_apple_technologies Ranked matches in the 405-technology snapshot
get_apple_technology Compact local routes by default; explicit full guide view
search_local_references Local guides and reusable source matches without bodies
get_reference_outline Heading offsets for focused section reads
read_local_reference Exact source/guide content with bounded output and continuation
get_apple_updates Search of 96 update/release-note landing pages
plan_ios_app Implementation workflow and CLI executable/argument array
plan_app_icon Editable foreground-layer specification and Icon Composer workflow

Start a local MCP connection:

npx -y --package=ios-agent-mcp@2.4.0 ios-agent-knowledge

For a remotely hosted ChatGPT MCP connection, run the HTTP transport behind your host’s HTTPS ingress:

npx -y --package=ios-agent-mcp@2.4.0 ios-agent-knowledge --http --host 0.0.0.0 --port 3000

The MCP route is /mcp; readiness is /health. Without --http, the process uses stdio. The default HTTP host is loopback. The Dockerfile at mcp-server/Dockerfile.knowledge builds this public-reference service from the repository root.

This service accepts only search strings, indexed reference IDs/ranges, catalog IDs and app/icon briefs. It reads a bundled allowlisted knowledge file, never arbitrary caller-supplied filesystem paths. It has no source upload, code execution, build or publishing tools, makes no runtime network requests, and does not store briefs. Configure HTTPS, request limits and operational logging on the deployment host. Account-specific or private-data tools would need a separate authentication design.

The data has explicit snapshot dates. Refresh source snapshots in the repository and rebuild before release. A successful HTTP handshake proves protocol compatibility, not that an arbitrary host or client account has been configured.

Read docs/tooling/offline-source-library.md for offline search, source reuse, attribution and verification limits.

Anti-Patterns

  • Exposing the local analyzer to the internet with arbitrary filesystem paths.
  • Advertising a localhost address as a published ChatGPT endpoint.
  • Treating plan_ios_app as an autonomous code generator; the coding agent performs implementation.
  • Describing catalog snapshots as live Apple updates.

Sources: OpenAI MCP server guidance, plugin submission. A public directory submission needs a stable HTTPS endpoint and publisher verification; a skills-only plugin bundle is also supported.