Expose un document OpenAPI sous la forme d'un catalogue d'outils orienté recherche, permettant aux agents IA de découvrir et d'invoquer des opérations d'API via un backend distant.
Aussi connu sous sezzlee/mcp, openapi, @sezzlee/openapi-mcp
- Étoiles
- 1
- Outils
- —
- Téléchargements, 30 jours
- —
- Dernière version
- 1.0.0
- Dernier commit
- 13 nov. 2026
- Licence
- MIT
Qu'est-ce que Sezzlee OpenAPI MCP ?
Sezzlee OpenAPI est un serveur MCP. Expose un document OpenAPI sous la forme d'un catalogue d'outils orienté recherche, permettant aux agents IA de découvrir et d'invoquer des opérations d'API via un backend distant.
Résumé généré à partir du README du projet
À propos du MCP server Sezzlee OpenAPI
Sezzlee OpenAPI agit comme une passerelle qui transforme un document OpenAPI existant en un catalogue d'outils interrogeable pour les agents IA. Il est conçu pour les développeurs qui doivent exposer des opérations de backend distantes à un client IA sans intégrer directement un SDK embarqué dans leur processus backend. Le serveur lit les spécifications Swagger 2.0 ou OpenAPI 3.0–3.2 à partir d'un chemin de fichier local ou d'une URL distante pour découvrir les points de terminaison disponibles.
Les utilisateurs interagissent avec le serveur via des outils spécifiques, notamment search_tools, load_tool, invoke_tool et read_result. Le serveur peut fonctionner via le transport stdio ou via streamable HTTP. Pour les configurations distantes, il prend en charge les identifiants de ressources protégées par OAuth et l'échange de jetons RFC 8693 pour gérer l'authentification. La configuration est gérée via un fichier JSON spécifié par la variable d'environnement SEZZLEE_OPENAPI_CONFIG, permettant un contrôle précis sur les noms des outils, les termes de recherche et la sélection des opérations.
Ce que vous pouvez faire avec
- 1Exposer des API distantes aux agents IAConnectez un client IA à un service backend qui ne possède pas de SDK intégré en utilisant son document OpenAPI.
- 2Gérer la visibilité et le nommage des outilsConfigurez les routes API spécifiques qui sont exposées à l'agent et remplacez leurs noms par défaut.
- 3Automatiser les mises à jour de la documentation APIConfigurez le serveur pour qu'il relise périodiquement la source OpenAPI afin de refléter les changements du backend.
Extrait du README du projet. Chaque cas d'utilisation est vérifié par rapport à une citation du document.
Comment installer Sezzlee OpenAPI MCP
Formats vérifiés le 10 nov. 2026Ajouter à CursorAjouter à GooseInstaller Sezzlee OpenAPI MCP dans Claude Code
$ claude mcp add --env 'SEZZLEE_OPENAPI_CONFIG=<SEZZLEE_OPENAPI_CONFIG>' openapi -- npx -y @sezzlee/openapi-mcpInstaller Sezzlee OpenAPI MCP dans Cursor
~/.cursor/mcp.json
{
"mcpServers": {
"openapi": {
"command": "npx",
"args": [
"-y",
"@sezzlee/openapi-mcp"
],
"env": {
"SEZZLEE_OPENAPI_CONFIG": "<SEZZLEE_OPENAPI_CONFIG>"
}
}
}
}Installer Sezzlee OpenAPI MCP dans VS Code
.vscode/mcp.json
{
"servers": {
"openapi": {
"type": "stdio",
"command": "npx",
"args": [
"-y",
"@sezzlee/openapi-mcp"
],
"env": {
"SEZZLEE_OPENAPI_CONFIG": "<SEZZLEE_OPENAPI_CONFIG>"
}
}
}
}Installer Sezzlee OpenAPI MCP dans VS Code CLI
$ code --add-mcp '{"name":"openapi","type":"stdio","command":"npx","args":["-y","@sezzlee/openapi-mcp"],"env":{"SEZZLEE_OPENAPI_CONFIG":"<SEZZLEE_OPENAPI_CONFIG>"}}'Installer Sezzlee OpenAPI MCP dans Claude Desktop
claude_desktop_config.json
{
"mcpServers": {
"openapi": {
"command": "npx",
"args": [
"-y",
"@sezzlee/openapi-mcp"
],
"env": {
"SEZZLEE_OPENAPI_CONFIG": "<SEZZLEE_OPENAPI_CONFIG>"
}
}
}
}Installer Sezzlee OpenAPI MCP dans Codex
$ codex mcp add openapi --env 'SEZZLEE_OPENAPI_CONFIG=<SEZZLEE_OPENAPI_CONFIG>' -- npx -y @sezzlee/openapi-mcpInstaller Sezzlee OpenAPI MCP dans Codex config.toml
~/.codex/config.toml
[mcp_servers.openapi] command = "npx" args = ["-y", "@sezzlee/openapi-mcp"] [mcp_servers.openapi.env] SEZZLEE_OPENAPI_CONFIG = "<SEZZLEE_OPENAPI_CONFIG>"
Installer Sezzlee OpenAPI MCP dans Gemini CLI
$ gemini mcp add -s user -e 'SEZZLEE_OPENAPI_CONFIG=<SEZZLEE_OPENAPI_CONFIG>' openapi npx -- -y @sezzlee/openapi-mcpInstaller Sezzlee OpenAPI MCP dans Gemini settings.json
~/.gemini/settings.json
{
"mcpServers": {
"openapi": {
"command": "npx",
"args": [
"-y",
"@sezzlee/openapi-mcp"
],
"env": {
"SEZZLEE_OPENAPI_CONFIG": "<SEZZLEE_OPENAPI_CONFIG>"
}
}
}
}Configuration
SEZZLEE_OPENAPI_CONFIGRequisPath to the server config file
| Nom | Description | Type |
|---|---|---|
| SEZZLEE_OPENAPI_CONFIG | Path to the server config file | Requis |
Vérifications automatisées
Exécuté le 14 nov. 2026Rapport complet →A passé nos vérifications automatisées. Il ne s'agit pas d'un audit de sécurité : lisez le code et donnez au serveur l'accès le plus restreint dont il a besoin. Fonctionnement des vérifications
- Dépôt source lié et accessiblesezzlee/mcp is public and reachablegithub.com/sezzlee/mcp ↗Réussi
- Le nom du package correspond au dépôt source@sezzlee/openapi-mcp names sezzlee/mcp as sourcenpmjs.com/package/@sezzlee/openapi-mcp ↗Réussi
- Provenance du build (attestation Sigstore)Sigstore attestation for @sezzlee/openapi-mcpnpmjs.com/package/@sezzlee/openapi-mcp/v/1.0.0#provenance ↗Réussi
- Licence open source (OSI)MITspdx.org/licenses/MIT.html ↗Réussi
- Commit ou version publiée au cours de l'année écouléeLast commit or release 2026-11-09github.com/sezzlee/mcp ↗Réussi
- OpenSSF ScorecardNon vérifié
- Nom du package non proche d'un package populaireNo popular package one edit away from @sezzlee/openapi-mcpRéussi
- Aucun avis critique dans le package ou ses dépendances10 resolved dependencies checked; no critical advisoryosv.dev ↗Réussi
- Aucune instruction cachée destinée au modèle0 tool descriptions and the README read; no override phrases, credential paths or hidden charactersRéussi
- Les outils destructifs sont étiquetésNon vérifié
- Le point de terminaison distant répond à la poignée de main MCPNon vérifié
README
Voir sur GitHub@sezzlee/openapi-mcp
MCP server that exposes an OpenAPI (Swagger 2.0 / OpenAPI 3.0–3.2) document as sezzlee's
search-first tool catalog — search_tools, load_tool, invoke_tool, read_result — over a remote
backend, calling it with fetch.
Status:
1.0.0, stable. Install withnpm install -g @sezzlee/openapi-mcpor run it withnpx -y @sezzlee/openapi-mcp; the binary issezzlee-openapi.
How this differs from the embedded SDK
The NestJS SDK sits inside your backend process: it discovers endpoints from your own controllers/routes and replay each MCP call through your existing pipeline, so your authentication and authorization run exactly as they do today.
openapi-mcp is a separate service. It has no access to a backend's process or pipeline — it
reads the backend's OpenAPI document to learn what operations exist (the ingestion
rules), and calls the backend over the network for every
invocation, attaching a credential it resolves itself (how a credential is chosen and
written). This is the only path for a backend that has no
embedded SDK integrated into it.
Everything an embedded SDK declares in code, a document declares in one x-sezzlee extension per
operation, and a host that cannot edit the document declares under operations in this config:
selection, tool name, prefix, behaviour hints, argument curation, variants, families, search terms
and revisions (the x-sezzlee extension).
A backend that routes many capabilities through one endpoint (POST /Rest/InvokeDynamicMethod/{methodId})
lists them in that operation's x-sezzlee.family; each member becomes its own tool, with its own
name, description and body, and the dispatch key stays hidden.
Quick start
The server reads its configuration from the file named by SEZZLEE_OPENAPI_CONFIG. Build the
package first (this repository never runs @sezzlee/core's consumers against a stale dist):
pnpm turbo run build --filter=@sezzlee/openapi-mcp
A minimal config — a local document, opt-in selection so at least one operation is exposed,
default stdio transport:
{
"source": "./openapi.json",
"selection": { "default": "include" }
}
SEZZLEE_OPENAPI_CONFIG=/absolute/path/to/config.json node packages/servers/openapi-mcp/dist/cli.js
source is either a path (resolved relative to the config file's own directory) or an
http(s):// URL. On startup the server ingests the document, builds the catalog, prints a
one-line-per-diagnostic-code summary to stderr, and then serves stdio or listens for
streamable HTTP, depending on transport.kind.
A config that fails validation, or that names an environment variable that is not set, stops the
process with a message on stderr and exit code 2. A document that ingests with a fatal
diagnostic, or a catalog with a fatal diagnostic (name_collision, invalid_name,
ambiguous_selection, …), stops it with exit code 1.
Configuration
All keys below are read from the SEZZLEE_OPENAPI_CONFIG JSON file and validated with a zod
schema (src/platform/config.ts) that rejects unknown keys.
| Key | Type | Default | Notes |
|---|---|---|---|
transport.kind |
"stdio" | "http" |
"stdio" |
— |
transport.host |
string (http only) |
"127.0.0.1" |
Must stay loopback (127.0.0.1, ::1, localhost) unless tokenExchange is configured — see below |
transport.port |
int, 0–65535 (http only) |
8787 |
— |
transport.path |
string starting with / (http only) |
"/mcp" |
— |
transport.resource |
url (http only) |
required | The OAuth protected-resource identifier advertised at /.well-known/oauth-protected-resource |
transport.authorizationServers |
array of url, min 1 (http only) |
required | — |
transport.allowedHostnames |
array of string (http only) |
[] |
Extra Host header values accepted, on top of the localhost defaults |
tokenExchange |
object | not configured | RFC 8693 token exchange; required for an http transport to authenticate a caller (see Rules) |
tokenExchange.tokenEndpoint |
url | required | — |
tokenExchange.clientId |
string | required | — |
tokenExchange.clientSecret |
{ fromEnv: string } |
required | — |
tokenExchange.clientAuth |
"basic" | "post" |
"basic" |
client_secret_basic or client_secret_post |
tokenExchange.audience |
string | not sent | — |
tokenExchange.resource |
url | not sent | — |
tokenExchange.scope |
string | not sent | — |
tokenExchange.schemes |
array of string, min 1 | required | The security scheme names the exchanged token satisfies |
source |
string, non-empty | required | A local file path (relative to the config file) or an http(s):// URL to the OpenAPI document |
reload.intervalSeconds |
integer ≥ 10 | not set | Re-reads source this often, plus up to 10% jitter. Conditional on the last ETag/Last-Modified for an http(s) source; a body identical to the served one is ignored |
reload.onSignal |
boolean | true |
SIGHUP re-reads source. Without a reload block the document is read once at startup and SIGHUP keeps its default action |
baseUrl |
url | not set | Replaces the document's root servers |
serverVariables |
record<string, string> | not set | Overrides for the document's server variable defaults |
hoistPathPrefix |
string starting with / |
not set | A leading path segment moved from every route into the base URL |
outputSchema |
"document" | "omit" |
"document" |
omit for a backend whose responses do not match its document |
requestBodyRequired |
"document" | "always" |
"document" |
always treats an undeclared requestBody.required as true |
strict |
boolean | false |
Raises openapi_document_invalid from a warning to a fatal diagnostic |
selection.default |
"include" | "exclude" |
"exclude" |
Whether an operation is exposed when no rule matches it (opt-in by default) |
selection.rules |
array of { route?, method?, decision } |
[] |
route/method are matched the same way the embedded SDKs match a controller route |
operations |
record<operation key, x-sezzlee object> |
{} |
Laid over each operation's x-sezzlee field by field; a key that names no operation is fatal. A field here may not also be set through names, searchTerms or revision.operations |
names |
record<operation key, string> | {} |
Overrides a tool's name; the value must match ^[a-z][a-z0-9_]{0,255}$ |
revision.catalog |
integer ≥ 1 | not configured | Behaviour revision of every tool; comes before the document root's x-sezzlee revision |
revision.paths |
record<document path, integer ≥ 1> | {} |
Behaviour revision of every operation under that path item; comes before the path item's x-sezzlee revision |
revision.operations |
record<operation key, integer ≥ 1> | {} |
Behaviour revision of one operation; comes before its x-sezzlee revision. The levels add up rather than replacing one another |
credentials |
record<security scheme name, credential> | {} |
Either { "value": { "fromEnv": "..." } } or { "username": { "fromEnv": "..." }, "password": { "fromEnv": "..." } }; see credentials.md |
allowHosts |
array of string | [] |
Hosts a backend operation may be served from, beyond the document's own root server |
refHosts |
array of string | [] |
Hosts an external $ref may be fetched from, beyond the document's own host. Kept apart from allowHosts, so allowing a schema host never lets calls or credentials go there |
identityCookies |
array of string | [] |
Cookie names treated as identity carriers on top of the default deny-list (compared case-insensitively); they extend it and never replace it |
limits.timeoutMs |
positive int | 30000 |
Per-invocation deadline |
limits.maxResponseBytes |
positive int | 262144 |
The largest answer that reaches the agent; a larger success is stored for read_result while results.enabled holds, and refused otherwise |
limits.maxBufferedBytes |
positive int | 16777216 |
The largest backend body read, applied while the body streams; a larger one is refused unread |
results.enabled |
boolean | true |
Keeps over-budget successes in memory and publishes read_result; see result-reading.md |
results.lifetimeMs |
int ≥ 1000 | 600000 |
How long a stored result is kept |
results.maxCallerBytes |
positive int | 16777216 |
What one caller's stored results may hold; the oldest go first. A caller is the bearer token it authenticated with, or every loopback caller when there is no token exchange |
results.maxBytes |
positive int | 67108864 |
What all stored results may hold; at least results.maxCallerBytes |
limits.maxInlineFileBytes |
positive int | 1048576 |
— |
A credentials entry is keyed by the security scheme name it satisfies; a tokenExchange.schemes
entry marks that same scheme as satisfied by the exchanged token instead. credentials values are
never inlined — only a reference to an environment variable — so a config file can be committed
and shared.
Rules
- Only
src/net/fetch.tsreaches the network. It is the one place the host allowlist, the manual redirect handling and the byte cap are applied;src/transport/http.tsmay usenode:httponly to listen, never to call out. Redirects are never followed — a followed redirect could carry a request, and its credential, to a host the allowlist never approved. - Only
src/platform/files.tstouches the filesystem. A file the document references through an external$ref, or named bysource, is checked against the directory the document was read from on its realpath, so neither a../segment nor a symbolic link can escape it. - The caller's MCP token is never forwarded to the backend. Its audience is this server, not
the backend; forwarding it would bypass the backend's own audience check. On the
httptransport withtokenExchangeconfigured, the caller's token is exchanged (RFC 8693) for a backend-scoped token at the transport's bearer gate, and only the exchanged token reachesinvoke_tool. - An
httptransport withouttokenExchangeauthenticates no caller, so it may only bind a loopback host (127.0.0.1,::1,localhost); otherwise the operator's static credentials would be reachable from the network. ConfiguringtokenExchangeon thestdiotransport is likewise refused (token_exchange_requires_http) — there is no caller token onstdioto exchange. - A reload never widens egress.
reloadre-reads only the document, never the config, and the hosts the first catalog may call are the ceiling for every later one: a document whose servers move to a host outside it, from https to plain http, or that sends a credential to a plain-http origin that did not carry it before, is refused (widens_egress). A refused document, one with a fatal diagnostic, or one that cannot be read keeps the running catalog. A committed reload advances the catalog generation stamped on every meta-tool and sendstools/list_changed. A tool loaded before a reload answersunknown_toolif it is gone andtool_changedto a call pinned to its old version. - Visibility is not enforcement: an operation's
securitysays which credential a call needs, not which caller may make it, so the catalog reports every operation's identity asunknownunless its document sayssecurity: [].
Development
pnpm turbo run build --filter=@sezzlee/openapi-mcp
pnpm turbo run lint --filter=@sezzlee/openapi-mcp
pnpm turbo run check-types --filter=@sezzlee/openapi-mcp
pnpm turbo run test --filter=@sezzlee/openapi-mcp
Run these through Turbo, not pnpm --filter @sezzlee/openapi-mcp <task> — the bare filter skips
dependsOn: ["build"] and the test task would run against a stale dist.
test/ holds unit suites that need no environment variables, plus one gated suite:
acceptance.spec.ts— skipped unlessSEZZLEE_OPENAPI_ACCEPTANCE_DOCnames a path to a real backend's OpenAPI document. The document itself never enters the repository; the suite ingests it, builds a catalog, asserts there is no fatal diagnostic, and prints the diagnostic summary instead of pinning a snapshot that would copy that backend's surface into the tree.SEZZLEE_OPENAPI_ACCEPTANCE_DOCis declared in this package'sturbo.jsontesttask so Turbo passes it through.
Questions sur Sezzlee OpenAPI MCP
Comment ajouter Sezzlee OpenAPI MCP à Claude Code ?
Exécutez claude mcp add --env 'SEZZLEE_OPENAPI_CONFIG=<SEZZLEE_OPENAPI_CONFIG>' openapi -- npx -y @sezzlee/openapi-mcp dans votre terminal. Ajoutez --scope user pour le rendre disponible dans tous les projets.
Comment ajouter Sezzlee OpenAPI MCP à Cursor ?
Ouvrez ~/.cursor/mcp.json et ajoutez le serveur sous mcpServers, avec la configuration Cursor indiquée dans la section d'installation ci-dessus.
Quelles méthodes de transport sont prises en charge ?
Le serveur prend en charge les transports stdio et streamable HTTP.
Puis-je utiliser un fichier OpenAPI local comme source ?
Oui, la source peut être soit un chemin de fichier local relatif au fichier de configuration, soit une URL HTTP(S).