activeing123/mcptoon v0.8.26

v0.8.26 — OpenAPI/GraphQL passthrough, OAuth, MCP resources/prompts/complete, saved connections

on GitHub

Added

• mcptoon openapi list|call — drive any OpenAPI / Swagger REST API as if it
  were MCP tools. One command per operation, name=value params, a cached spec
  keyed by a caller-supplied --cache-key (traversal-safe), --cache-ttl, and
  --base-url to point the same spec at prod or staging. Reuses the shared safety
  gates (dangerous-op, poisoning, credential-leak) and usage tracking.
• mcptoon graphql list|call — the same shape for GraphQL endpoints:
  list enumerates queries/mutations, call runs one with --fields "a b".
• mcptoon auth login|creds|status — OAuth for protected APIs. login runs
  the PKCE browser flow, creds the client-credentials grant, status reports
  token state (never the token). --register URL performs RFC 7591 Dynamic Client
  Registration so no client-id has to be pasted; --client-name sets the name the
  server records.
• mcptoon resources list|read, mcptoon prompts list|get, mcptoon complete —
  the MCP primitives beyond tools: read a server's resources, render its prompt
  templates, and ask it to complete an argument value
  (<ref>:<arg>=<prefix>, split on the first = after the arg name).
• mcptoon bake create|list|show|update|remove|install|run — save a connection
  under a short name and reuse it. Literal secrets are rejected; headers accept
  only env:VAR or file:PATH references, and install writes a launcher that
  carries only the bake name, never a URL or key.
• manifest --sort alpha|usage|recent, manifest --top N, --verbose — order
  the tool list by name / your own call frequency / recency, cap it to N tools per
  server, and optionally add descriptions.
• --transport auto|sse|streamable — process-wide preference for how HTTP MCP
  servers are negotiated (a per-server http_transport in config still wins).

Changed

• README restructured. The hero is now three cards matching the three jobs a
  user actually has — save context (tools, skills and docs), configure every
  agent once (a native CLI: any shell-capable agent gets every feature; sync
  covers the GUI clients), and find the right tool among 17,000+ (discover picks
  up what is already installed, so nothing is reinstalled). The capability table is
  regrouped to match those three, and now names the API-passthrough, OAuth, MCP
  resource/prompt and bake features that earlier versions omitted.

Fixed

• Stale skill-catalog figures. The READMEs and the two benchmark docs still
  quoted the 371-skill era (926,232 tokens, 501-token lookup, 7.07
  context windows). Measured on this machine with tiktoken cl100k_base over all
  six skill roots, de-duplicated by real path: 480 skills, 1,572,330 tokens,
  791-token lookup, 12.00 windows. All surfaces now carry the current numbers and
  the caliber that produces them.
• Stale wheel size. Every surface said 363KB; the LF-built wheel for this
  tree is 534KB (547,133 bytes). Updated on all eight claim surfaces and the
  footprint guard now pins 534KB (363KB retired).