Openai Docs

OpenAI Docs

• MUST use MCPorter openai-docs for official documentation
• Public server: NEVER authenticate
• MUST ground answers in fetched documentation
• NEVER invent undocumented behavior

Call and recover

• The known recipes below SHOULD be called directly without inventory or schema discovery
• Live server schema MUST remain authoritative when inspected
• Unknown optional arguments or input-validation failures: MUST inspect only targeted live schema, then retry once:

mcporter list openai-docs.<tool> --schema

• Output fields: MUST inspect actual results
• Tool declarations MAY omit output schemas

Use these live tools:

mcporter call openai-docs.search_openai_docs query='responses api tools'
mcporter call openai-docs.fetch_openai_doc url='https://developers.openai.com/...'
mcporter call openai-docs.list_openai_docs
mcporter call openai-docs.list_api_endpoints
mcporter call openai-docs.get_openapi_spec url='https://developers.openai.com/...'

Complex arguments SHOULD use --args '<JSON object>'.

Routes

Documentation

1. Compact query (2-6 terms): MUST use search_openai_docs
2. MUST fetch the best URL; anchor requires a known section
3. MUST cite the narrow source before broadening search

• Clear query absent: MAY browse with list_openai_docs
• MCP unavailable/unhelpful: MUST use official OpenAI domains
• MUST cite the fetched page

API reference

• Endpoint discovery MUST use list_api_endpoints
• Endpoint schemas MUST use get_openapi_spec
• MUST pair schemas with relevant guides or references
• Optional output controls: MUST inspect targeted live schema first

API troubleshooting

• You MUST first distinguish pre-response DNS, TLS, or network failures from API responses
• You MUST classify 401 from the actual error payload and headers as authentication
• You MUST classify 403 from the actual error payload and headers as project, model, or permission access
• For 429, you MUST use the actual error payload and headers to distinguish insufficient_quota from rate limiting
• You MUST use current official documentation via the documentation route for remediation
• NEVER guess about configuration errors or blindly retry them

Models and Codex

• Latest/current/default model: MUST fetch https://developers.openai.com/api/docs/guides/latest-model.md first
• Explicit model targets MUST win. NEVER migrate silently

Unspecified migration or prompt upgrade: MUST run the resolver:

uv run --script <skill-dir>/scripts/cli.py latest-model

• Resolver output MUST include all three fields: model, migrationGuideUrl, promptingGuideUrl.
• MUST fetch returned guides through fetch_openai_doc

Broad Codex self-knowledge: MUST run the helper in a writable session:

uv run --script <skill-dir>/scripts/cli.py codex-manual

• MUST read only relevant outline/manual sections
• Helper unavailable/insufficient: MUST use the documentation route
• Official evidence absent: MUST state bounded uncertainty

Required follow-up reads