mcp-dev
/mcp-dev
Projektuje, buduje i wysyła serwery MCP i aplikacje MCP dla Claude oraz ChatGPT, włącznie z OAuth i recenzją katalogu.
- Autor
- travnie ↗
- Repozytorium
- twojstar ↗
- Licencja
- ISC
- Ostatnia zmiana repozytorium
- 13.11.2026
Ocena i bezpieczeństwo
- Pliki w paczce
- 16
- Skrypty
- 3
- Komentarze HTML
- 0
- Adresy URL w SKILL.md
- 1
Brak sygnałów. Repozytorium zmienione 13.11.2026, bez odwołań do wycofanych modeli i API.
ISC, otwarta.
- 3 skrypty w paczce. Przejrzyj je przed instalacją, bo Claude może je uruchomić.
Skrypty: scripts/eval/connections.py, scripts/eval/evaluation.py, scripts/mcp_probe.py
Co robi i kiedy
Projektuje, buduje i wysyła serwery MCP i aplikacje MCP dla Claude oraz ChatGPT, włącznie z OAuth i recenzją katalogu.
Wyzwalacze z opisu
Oryginalny opis (EN)
Design, build, test and ship MCP servers and MCP Apps for Claude and ChatGPT/Codex: deployment choice, tool design, OAuth (CIMD/DCR), MCPB, directory review, spec changes. Use for any 'build/fix/review an MCP server or connector'.
Jak działa
- 1. Pick the shape (ask only what the brief doesn't answer)
- 2. Tools
- 3. Auth
- 4. Test
- 5. Ship
- Chat vs Claude Code
Podgląd SKILL.md
---
name: "mcp-dev"
description: "Design, build, test and ship MCP servers and MCP Apps for Claude and ChatGPT/Codex: deployment choice, tool design, OAuth (CIMD/DCR), MCPB, directory review, spec changes. Use for any 'build/fix/review an MCP server or connector'."
license: Apache-2.0
---
# MCP servers + apps
The spec moved to **2026-07-28** (stateless: no `initialize` handshake, no `Mcp-Session-Id`, `server/discover`, elicitation/sampling via multi-round-trip results, `resultType` on every result). Hosts and SDKs migrate at different speeds — check `mcp-docs` (`/specification/<rev>/changelog.mdx`) and the SDK changelog before writing protocol-level code, and let the official SDK negotiate versions. Platform rules: `claude-ai-docs` (`/connectors/building/*`), `openai-docs` (`/plugins/build/mcp-server`, `/plugins/reference`).
## 1. Pick the shape (ask only what the brief doesn't answer)
| Connects to | Users | Shape |
|---|---|---|
| Cloud API / SaaS / pure logic | anyone | **Remote streamable HTTP** (default). Works in chat, Cowork, Code, ChatGPT |
| Local files, desktop app, localhost, OS | installers | **MCPB** bundle (runtime included) — Claude desktop/Code only; chat ignores local servers |
| Just me, prototype | me | local stdio (`npx`/`uvx`), upgrade path MCPB |
| Needs pickers, charts, previews in chat | any | **MCP App** = remote server + `ui://` resources → `references/apps.md` |
| Ships workflows/know-how with the tools | any | **Skills over MCP** (`io.modelcontextprotocol/skills`, SEP-2640) → `references/skills.md` |
Bartek default: Cloudflare Worker (`cloudflare` skill): new servers use `createMcpHandler` from `agents/mcp/server` with an SDK v2 server factory (stateless, no Durable Object); `McpAgent` is deprecated and feature-frozen since Agents SDK 0.20, only for migrating old servers. State lives in KV/D1/DO keyed by user, not by session. Service bindings for same-account Worker calls. Portable: official TS SDK v2 = split packages `@modelcontextprotocol/server` (+ `client`, `node`, `express`); `@modelcontextprotocol/sdk` is the v1 line: new code on v2 (docs index `ts.sdk.modelcontextprotocol.io/v2/llms.txt`); migrate v1 projects with `npx @modelcontextprotocol/codemod@latest v1-to-v2 .`, and until then each module uses one API. Python: FastMCP (jlowin's 3.x, not the frozen 1.0 in `mcp`). Versions: use the exact pin a framework's docs require (Cloudflare Agents pins `@modelcontextprotocol/server`); otherwise `npm install <pkg>`, never versions typed from memory.
TS SDK v2 traps:
- `zod` ^4.2.0 with `import * as z from "zod/v4"`: Zod 3 is unsupported, 4.0–4.1 fail type-checking.
- `server.tool/resource/prompt` are removed: use `registerTool/registerResource/registerPrompt`.
- Stateless HTTP: a new `McpServer` and transport per request from the factory, never one module-scope instance. Stateful sessions keep one per session.
- Node.js 20+ (v1 ran on 18).
## 2. Tools
Wrapping a REST/OpenAPI API: curate, don't mirror every endpoint (`references/tools.md` → OpenAPI). ≤15 actions → one tool each. 30+ → `search_actions` + `execute_action`, promote the top 3-5. Details and directory rules: `references/tools.md`. Non-negotiable for Claude's directory: read and write in separate tools, `title` + `readOnlyHint`/`destructiveHint` on every tool, names ≤64 chars, descriptions describe (never instruct the model), actionable errors.
## 3. Auth
Prefer: none → OAuth with CIMD (AS metadata advertises `client_id_metadata_document_supported: true` **and** `none` in `token_endpoint_auth_methods_supported`, else Claude falls back to DCR) → DCR. Callback `https://claude.ai/api/mcp/auth_callback`; Claude Code loopback needs port-agnostic `localhost` + `127.0.0.1`. No `client_credentials`; static header credentials are org-owner beta only. Validate token audience (RFC 8707); never pass the user's token upstream. ChatGPT side declares auth in portable `mcp.json` (`extensions.com.openai.auth`). More: `references/ship.md`.
## 4. Test
1. `python3 scripts/mcp_probe.py <url>` — detects spec era (`server/discover` vs `initialize`), lists tools with annotations, flags directory-rule violations; `--call name '{json}'` runs one tool. Works from the chat sandbox for public URLs.
2. MCP Inspector (`npx @modelcontextprotocol/inspector`) in Claude Code.
3. Real host: claude.ai Settings → Connectors → custom connector (Claude sends `clientInfo.name: "claude-ai"`); ChatGPT Plugins → Add custom MCP server; Grok Build `grok mcp add --transport http <n> <url>` + `grok mcp doctor <n>`. Local servers via tunnel.
4. Exercise every tool with valid and invalid input; check response size.
5. Quality, not just liveness: 10-question eval → `references/evals.md`.
Report the level reached — static review → syntax/build → server answers (`mcp_probe.py`) → real host loop — and what didn't run. "Files exist" is not "works".
## 5. Ship
Wrap the server in a plugin with a skill (`plugin-dev`) — the usual partner shape. Directory submission checklists: `references/ship.md`.
## Chat vs Claude Code
Chat sandbox: write code, run `mcp_probe.py` against deployed URLs, deploy via CI (house rule: workflow beats local). Claude Code: run servers locally, Inspector, `wrangler dev`, MCPB packing (`npx @anthropic-ai/mcpb pack`).
Pliki w paczce (16)
- LICENSE.txt 11 KB
- NOTICE.md 2 KB
- SKILL.md 5 KB
- agents/openai.yaml 1 KB
- assets/icon.svg 1 KB
- references/apps.md 8 KB
- references/evals.md 2 KB
- references/ship.md 3 KB
- references/skills.md 3 KB
- references/tools.md 5 KB
- scripts/mcp_probe.py 8 KB
- scripts/eval/LICENSE.txt 11 KB
- scripts/eval/connections.py 5 KB
- scripts/eval/evaluation.py 13 KB
- scripts/eval/example_evaluation.xml 1 KB
- scripts/eval/requirements.txt 1 KB
Wersje
| Repozytorium | Wynik | Licencja |
|---|---|---|
| Ta strona · travnie/twojstar | 90 | ISC |
| xiaolai/vmark | 90 | ISC |
Migawka bazy z 13.11.2026 · wersja paczki e0844ae467fe (skrót SHA-256 całej paczki)
Pobrania w ostatnich 30 dniach: 0 · ranking
Zgłoś problem z tym skillem: szkodliwa treść, błędny opis albo prośba autora o usunięcie.
Podobne skille
- 100mcp-server-development
Projektuje i wdraża serwery MCP oraz narzędzia dla agentów: projekt narzędzi, zmiany specyfikacji, OAuth, bezpieczeństwo i wybór SDK, także przy migracji serwera. - 90publish-mcp-server
Prowadzi serwer MCP przez zgłoszenie do katalogów ChatGPT i Claude, od konfiguracji po uwagi recenzentów. - 95build-mcp-use-server
Buduje, poprawia i weryfikuje serwery MCP w TypeScript z mcp-use v2 oraz aplikacje ChatGPT z widokami i interfejsem. - 90build-mcp-apps
Prowadzi przez tworzenie i aktualizację serwerów MCP oraz aplikacji ChatGPT we frameworku Skybridge, od szkicu po wdrożenie. - 100chatgpt-app-builder
Buduje, zmienia, debuguje, migruje i przegląda serwery MCP w TypeScripcie oraz aplikacje MCP Apps z mcp-use: narzędzia, zasoby, prompty, middleware, widoki, uwierzytelnianie i skille nad MCP. - 100mcp
Integruje json-render z MCP Apps, aby serwery MCP renderowały interaktywne interfejsy w Claude, ChatGPT, Cursor lub VS Code.