MCP server de mcpify

mcpify

English | Türkçe

Turn any OpenAPI REST API into an MCP server — so Claude Code, Cursor, and every other MCP client can call your API directly. One command, zero runtime dependencies:

mcpify serve https://your-company.com/openapi.json

# try it right now, nothing installed (uvx pulls from PyPI on demand)
uvx --from mcpify-openapi mcpify list examples/petstore.json --cost

Focused, production-ready, CLI-first: one job (OpenAPI → MCP). Everything else — governance, credentials, token economics, operations — is opt-in and stays out of the way until you need it:

• Safe for agents by default — lazy context budget (--lazy), secret masking (--redact), blast-radius limits (--read-only, per-token RBAC)
• Ready for real deployments — stdio + HTTP/SSE, OAuth2 and split read/write credentials, health probes, audit trail, metrics, multi-API serving, hot reload
• Proven, not promised — 520 tests across twenty-eight suites, a hostile-spec corpus, live-CI checks, zero runtime dependencies

Why you'll like it

From spec to server

• 60 seconds to working — point it at any OpenAPI 3.x spec (file or URL)
• Every operation becomes a first-class MCP tool — input schemas are generated from parameters + requestBody, internal $refs are resolved
• Spec versions diffed from the tool view — mcpify diff old.yaml new.yaml reports added/removed/changed operations with per-change breaking verdicts and a migration guide; --fail-on-breaking is a CI gate
• mcpify doctor — tells you if your spec is agent-friendly before you ship: missing operationIds, missing summaries, instruction-like tool text, overlong descriptions; --probe dials the API once — with your real credential (--auth-env) when you want auth proven end-to-end, and --fail-on-http-error for a strict CI gate
• mcpify try / mcpify mock / mcpify output-server — call the tools without an agent client, serve a schema-shaped fake API for CI, or bake a serve command into a shareable script

Credentials & policy

• Credentials never touch the spec or the model — pulled from your environment at call time; the spec's own security declarations pick the flags (bearer, basic, header, query)
• OAuth2 client-credentials built in — tokens are fetched, cached, refreshed and re-fetched on a mid-flight 401 (RFC 6749, stdlib only); --write-oauth2-* gives non-GET calls a second client identity so reads and writes authenticate as different clients
• Least-privilege by default — --write-auth-env splits the static credential (reads on your read key, writes on a dedicated key), --read-only filters the surface, --deny/--allow hides mutating GETs, per-token RBAC gives each bearer token its own allow/deny scopes
• --redact password,token — values whose key names a secret are masked with *** at every level of every response (error bodies included, case-insensitive); the model never sees them
• Audit trail without content exposure — one JSON line per call: tool, API, status, latency, an argument fingerprint (never raw arguments); --plugin loads your Python module for auth/request/result hooks

Token economics

• See the bill before serving — mcpify list --cost prices the surface (~4 chars/token): what every agent pays in EVERY tools/list; multi-API configs get per-API and total prices in one run
• --fields id,event — response projection that selects at every level: selected keys keep their value, non-selected containers stay transparent, emptied containers drop. Live weather.gov: 350 alerts in full inside the budget that previously truncated at ~189
• Valid truncation — oversized responses are cut along JSON structure with an explicit "truncated": true marker, never mid-document
• --lazy search-then-call — cut api.weather.gov's listing by 95.5%; search results now show what pulling each full schema would cost, so the agent pulls only what it needs
• [tool-text] overrides — doctor flags model-facing instruction-like descriptions; you replace them per tool in config

Operations

• Two transports, one tool surface — stdio for local agents; serve --http 8080 speaks MCP Streamable HTTP (SSE responses for clients that ask, JSON otherwise) so a whole team shares one server, with optional bearer tokens
• Several APIs, one MCP server — [apis.NAME] sections in .mcpify.toml: per-API auth, caching, retries, filters and rate limits; collision-safe renames; aggregated health; mcpify status probes every API in parallel
• Upstream courtesy built in — ETag-aware caching, idempotent-only retries (502/503/504), --wait-on-429 honors Retry-After, --rate-limit RPS caps requests/second (per upstream in multi-API, retries included)
• Observability, opt-in only — Prometheus --metrics (call counters, latencies, cache, health — plus projection/redaction counters when those run), --otel spans, --reload hot swap, mcpify ui local dashboard
• Host it yourself for free — docker-compose with automatic-HTTPS Caddy plus a hardened systemd unit: Self-hosting guide
• Zero runtime dependencies — the entire tree is auditable stdlib Python; YAML specs need an optional pip install 'mcpify[yaml]'

Quick start

# install (installs the `mcpify` command)
pipx install mcpify-openapi

# run without installing (uvx — pulls from PyPI on demand)
uvx --from mcpify-openapi mcpify list ./openapi.json --read-only

# first time? the wizard writes a config for you
uvx --from mcpify-openapi mcpify init

# ...as a container (GHCR, published on every release)
docker run -i ghcr.io/furkan708/mcpify:latest serve ./openapi.json --read-only

# ...or from source
git clone https://github.com/furkan708/mcpify.git
cd mcpify && pip install .

# 1. preview the tools that will be generated (add --cost for the context price)
mcpify list examples/petstore.json

# 2. validate the spec is agent-friendly (+ --probe for a live pre-flight)
mcpify doctor examples/petstore.json

# 3. serve it over MCP
mcpify serve examples/petstore.json --base-url https://petstore.example.com/v1

# 4. no agent client at hand? try the tools in your terminal
mcpify try examples/petstore.json --base-url https://petstore.example.com/v1

# 5. or share it over HTTP with the whole team
mcpify serve examples/petstore.json --http 8080 --http-token $SHARED_TOKEN

With authentication

# Bearer token read from the environment (never hardcoded)
export PETSTORE_KEY="sk-..."
mcpify serve petstore.json \
  --base-url https://petstore.example.com/v1 \
  --auth-env PETSTORE_KEY \
  --auth-style bearer \            # optional: auto-detected from the spec
  --read-only

No explicit style needed in the common case — the spec's security declarations pick bearer/basic/header/query (with the right name) for you. For HTTP Basic, the env variable holds username:password: --auth-style basic --auth-env CREDS.

Flag	Meaning
--auth-env VAR	environment variable holding the credential
--auth-style bearer|basic|header|query	how it is sent (default: auto-detected from the spec)
--auth-name NAME	header / query name for non-bearer styles (e.g. X-API-Key)

With OAuth2 (client credentials)

For APIs behind an OAuth2 identity provider (RFC 6749 §4.4). Credentials live in the environment; the access token is fetched, cached until its expires_in, refreshed transparently, and re-fetched automatically once if the API answers 401 mid-flight:

export OAUTH2_CLIENT_ID="..."
export OAUTH2_CLIENT_SECRET="..."
mcpify serve api.json \
  --oauth2-token-url https://idp.example.com/oauth2/token \
  --oauth2-client-id-env OAUTH2_CLIENT_ID \
  --oauth2-client-secret-env OAUTH2_CLIENT_SECRET \
  --oauth2-scope "read write"        # optional; --oauth2-client-auth body for token endpoints that reject Basic

Split write identities too: --write-oauth2-token-url (+ client/scope flags) runs a second client-credentials flow for non-GET calls — reads authenticate as the read client, writes as the write client, each with its own token cache and the same 401 self-heal. Mutually exclusive with --write-auth-env (pick one credential kind for writes).

Multiple APIs in one server

Put several OpenAPI documents in one config and serve them as a single tool surface — no gateway, no per-API process:

# .mcpify.toml
[apis.catalog]
spec = "https://shop.example.com/openapi.json"
auth-env = "CATALOG_TOKEN"          # per-API credential
cache-ttl = 60
rate-limit = 5                      # per-API courtesy throttle (req/s)
redact = "password,client_secret"   # per-API response masking

[apis.crm]
spec = "./crm.yaml"
read-only = true                    # per-API policy
base-url = "https://crm.internal/v2"
fields = "id,name"                  # per-API response projection

[apis.weather]
spec = "https://api.weather.gov/openapi.json"
timeout = 10

Surface switches (--lazy, --enable-preview, --http, --format) are server-wide flags; credentials, policies, caching and retries are per-API.

mcpify list --cost      # preview every API and price each surface
mcpify serve            # stdio, all three APIs, prefixed on collisions
mcpify serve --http 8080
mcpify try              # REPL across every API
mcpify status           # probes each API concurrently

mcpify status reports per API — [catalog] reachable (status 200, 0.03s) — https://shop.example.com — 31 tools — and exits non-zero if any API is unreachable. When two APIs expose the same tool name (list_pets), both get renamed with their label (catalog_list_pets, crm_list_pets) so nothing silently wins; non-conflicting names stay untouched. The mcpify_health tool returns one report covering every API. Precedence per key: CLI flags > [apis.NAME] > [serve]. Pass a positional spec or [apis.*] sections — never both.

Plug it into your agent

Claude Code:

claude mcp add my-api -- mcpify serve openapi.json --read-only

Claude Desktop / Cursor / any MCP client (claude_desktop_config.json):

{
  "mcpServers": {
    "petstore": {
      "command": "mcpify",
      "args": ["serve", "~/specs/petstore.json", "--auth-env", "PETSTORE_KEY"]
    }
  }
}

HTTP transport (team-shared server) — run mcpify serve api.json --http 0.0.0.0:8080 --http-token $TOKEN once, then point HTTP-capable clients at it:

{
  "mcpServers": {
    "petstore": {
      "type": "http",
      "url": "http://your-host:8080",
      "headers": { "Authorization": "Bearer <token>" }
    }
  }
}

Now ask your agent: "list the pets, then create one named Milo" — it discovers list_pets and create_pet, fills the arguments, and performs real HTTP calls.

How operations become tools

OpenAPI	mcpify
operationId	tool name (sanitized; falls back to method_path)
summary / description	tool description the agent reads
deprecated: true	shown by mcpify list before you expose old endpoints
parameters (path/query/header)	individual typed arguments with enums
requestBody (JSON)	a body object argument
$ref pointers	resolved inline (components → real schemas)
servers[0].url	default base URL (override: --base-url)

The agent only ever sees the tool list and your API's JSON responses — mcpify adds no middleware, caches nothing you did not ask for, and sends credentials nowhere except your API.

Doctor

$ mcpify doctor my-api.json
openapi: 3.0.3
title:   Acme API
paths:   23
tools:   41 operations
servers: https://api.acme.com
warning: 12/41 operations have no operationId (names fall back to method_path)
warning: 30/41 operations have no summary (agents see no description)

Add --probe for a live pre-flight — after the static report, mcpify dials one argument-free GET (or the base URL) and reports reachability; a connection failure exits non-zero so CI and shell scripts stop before serving:

$ mcpify doctor https://api.weather.gov/openapi.json --probe
...
probe:    GET /alerts → 200 reachable (2.80s)

CLI reference

mcpify list <spec> [--tag T] [--include P] [--exclude P] [--read-only] [--json]
mcpify list --cost                        # price the surface (~4 chars/token)
mcpify list --cost --lazy                 # ...and the 3-meta-tool lazy surface
mcpify list --config .mcpify.toml --cost  # multi-API: every surface priced
mcpify serve <spec> [--base-url URL] [--server INDEX|NAME] [--name N] [--auth-env VAR]
                    [--auth-style bearer|header|query] [--auth-name NAME]
                    [--oauth2-token-url URL --oauth2-client-id-env VAR
                     --oauth2-client-secret-env VAR] [--timeout S]
                    [--read-only] [--tag T] [--include P] [--exclude P]
                    [--http [HOST:]PORT] [--http-token TOKEN] [--wait-on-429 SEC]
mcpify try <spec> [same serve flags]        # interactive REPL, no agent needed
mcpify output-server <spec> -o FILE [-- <any serve flags>]
mcpify ui <spec> [same serve flags]         # local dashboard (tool explorer, health, config)
mcpify mock <spec> [--http 8000] [--delay-ms N]
mcpify diff OLD NEW [--json] [--fail-on-breaking]        # upgrade report + CI gate
mcpify diff OLD NEW --probe [--auth-env V]               # ...+ live check of the NEW API + cost delta
mcpify config-schema                        # JSON Schema for .mcpify.toml (editor wiring)
mcpify doctor <spec> [--probe --auth-env V --fail-on-http-error]   # static audit + live pre-flight / CI gate

# token economics: --fields id,name (projection), --redact password,token (masking),
#   --rate-limit RPS (upstream courtesy, retries included)
# credential split: --write-auth-env WRITE_KEY_ENV or --write-oauth2-token-url (reads keep --auth-env)
# tool-text overrides: [tool-text.TOOL] description = "..." in .mcpify.toml

# multi-API: define [apis.NAME] sections in .mcpify.toml, then run
#   mcpify list|serve|try|status|ui (no positional spec) — one process, every API
# ops add-ons for serve/ui: --metrics [HOST:]PORT   --reload   --cache-warm
#   --audit-log FILE   --http-token-file FILE   --plugin FILE (repeatable)   --otel [ENDPOINT]

Notes & limitations

• JSON specs work out of the box; YAML specs need pip install 'mcpify[yaml]'
• External $ref targets (files or URLs) are bundled automatically at load; circular cross-file refs are left in place rather than unwound (surface skips what it cannot resolve)
• Oversized JSON responses truncate to valid JSON (first items + a truncation marker); non-JSON bodies cut at a character boundary
• --fields selects at every level by documented rule (selected keys verbatim, non-selected containers transparent); it is a projection, not a security boundary — use --redact when a field must never reach the model
• Request bodies are exposed as a single body object argument — predictable over clever
• HTTP transport serves one JSON-RPC response per request (batching was removed from the MCP spec); clients that send Accept: text/event-stream get it framed as a single SSE message event. Server-initiated streams (a GET stream with sessions) stay deliberately out of scope for a stateless server
• Spec versions: OpenAPI 3.x and Swagger 2.x roots are accepted; 3.x is the happy path

Hardened against the real world

mcpify is audited on every release against a 10-category checklist of MCP best practices and published production failure modes — not just our own examples:

• Hostile-spec corpus (12/12): circular $refs, multipart uploads, allOf schemas, server URL variables, relative base URLs, oversized responses — every scenario derived from a documented real-world failure, fixed, and locked in by a regression test. Sources include the arXiv study of REST→MCP generation across 18 real APIs.
• Live integration: the real api.weather.gov spec loads in CI — and the live checks found the last two real bugs (nested-truncation envelope splits, top-level-only projection) before any user did.
• MCP lifecycle enforced: tools are unreachable until the client completes the initialize handshake.
• Session survives upstream failures: a slow upstream (read timeout), a mid-response disconnect, or even an unexpected exception inside one tool becomes a clean isError tool result with remediation — never a dead stdio connection (the failure mode reported against other OpenAPI-to-MCP servers, replayed here and locked in by regression tests).
• Blast-radius controls: read-only mode, deny/allow policy layer, per-token RBAC, response projection + secret masking, rate limiting, 40k-char valid truncation, --timeout, credentials never logged.

Full checklist with per-item status: docs/AUDIT-CHECKLIST.md