convez.ai
MCP ServerIntegração & APIsAvançado
liangfb

SaaS to Agent (MCP Server)

liangfb/saastoagent

Converta APIs REST em agentes de IA com MCP + playground interativo

Ver no GitHub
BomPluginScore · de 0 a 100

Transforma APIs REST de SaaS/enterprise em ferramentas chamáveis por agentes de IA via MCP. Ingere especificações OpenAPI, enriquece semanticamente cada endpoint, gera MCP Servers em containers isolados, e oferece um playground com agente ReAct para interação em linguagem natural.

Instalação

$npx convez-ai liangfb/saastoagent

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
42
Forks
0
Acompanhando
0
Saúde
71%
Último commit
há 2 meses
Versão
—
Issues abertas
0
Contribuidores
1

Casos de uso

  • Integrar múltiplas APIs SaaS em um agente único sem código manual
  • Criar interfaces de linguagem natural para APIs legadas de enterprise
  • Automatizar orquestração de chamadas de API com raciocínio de agente

Veredito

Plataforma de referência sólida para transformar OpenAPI specs em ferramentas de agentes MCP. Arquitetura K8s bem estruturada, mas ainda em fase de desenvolvimento — alguns recursos (streaming de logs, anotação de traces) estão como stubs. Ideal para times que precisam expor APIs legadas a agentes de IA de forma governada.

README

A tradução para o português deste README ainda está na fila.

Agentic Service Mesh

Turn existing SaaS / enterprise OpenAPI (REST) endpoints into AI‑Agent‑callable MCP tools — then talk to them in natural language.

Upload (or paste) an OpenAPI/Swagger spec; the platform parses it, semantically enriches each endpoint, generates an MCP Server that runs as its own container, and lets a single ReAct Agent plan and call those tools end‑to‑end from a chat Playground.

Status: developer platform / reference implementation. The core pipeline (OpenAPI → semantic enhancement → MCP runtime → Agent) works end‑to‑end. Enterprise governance (OBO identity, CAEP, fine‑grained RBAC, encryption‑at‑rest) is on the roadmap, not shipped — see Security.

What it does (core)

  1. OpenAPI ingestion — import a spec by URL or by pasting JSON/YAML (for auth‑gated or disabled doc endpoints).
  2. Semantic enhancement — an LLM rewrites terse API metadata into Agent‑friendly tool descriptions.
  3. Dynamic MCP runtime — each source is deployed as an independent MCP Server (K8s pod) that proxies calls to the upstream API, injecting the configured credential (API key / Bearer / OAuth).
  4. Agent Playground — a single ReAct Agent, bound to a chosen tool set, plans and invokes tools; full execution traces stream live.
  5. Identity & memory — platform login (JWT), per‑user long‑term memory via mem0, and managed upstream API credentials.

Not yet implemented

As a reference implementation, a few observability surfaces are stubbed:

  • Live log streaming (GET /api/v1/logs/stream, SSE) returns an empty stream — log listing and filtering work; only the realtime push is a stub.
  • Trace annotation (POST /api/v1/traces/:id/annotate) returns a "pending" error — Agent traces are still recorded/queried via Langfuse.

Architecture

PathRole
backend/NestJS API server (port 8000, prefix /api/v1)
frontend/React 19 + Vite SPA (port 3000)
mcp-runtime/Generic MCP Server container image
mem0-server/Python mem0 service (Agent memory, OpenSearch‑backed)
mock-erp/5 Python FastAPI services for E2E testing
infra/Terraform + K8s manifests + deploy scripts

A single ReAct Agent (Mastra + Vercel AI SDK) replaces the earlier Router/Specialist multi‑agent design — one Agent reasons over its bound MCP tools directly.

Deployment (AWS / EKS)

The platform runs on AWS EKS. There are two distinct paths — bootstrap once, then iterate.

First-time deploy (bootstrap)

Creates all base resources (namespace, ConfigMap, Secret, Deployments, Service, Ingress).

# 1) Provision infrastructure (EKS, RDS, ECR, …)
cd infra && terraform init && terraform apply && cd ..

# 2) Create config from templates, then fill in your values.
cp infra/.env.example infra/.env                                   # AWS region / cluster / account
for ex in infra/k8s/*.yaml.example; do cp "$ex" "${ex%.example}"; done
# Edit infra/.env and the generated infra/k8s/*.yaml — especially
# infra/k8s/02-secrets.yaml (DATABASE_URL, JWT_SECRET_KEY, API keys, …).

# 3) Build + push images, then apply all manifests to create the resources.
./infra/scripts/build-and-push.sh
./infra/scripts/deploy-k8s.sh

Ongoing updates

Once the base resources exist, push new code with the rolling-update script (build → ECR → migrate → kubectl set image). It does not create resources and will abort if the namespace/deployments are missing — run the first-time deploy above for a fresh cluster.

./infra/scripts/deploy-test-env.sh                  # all services (also builds + pushes mcp-runtime)
./infra/scripts/deploy-test-env.sh --backend-only   # or a single service
./infra/scripts/deploy-test-env.sh --mcp-runtime-only  # rebuild the MCP runtime image

mcp-runtime has no standing deployment — the backend launches it on demand. After updating it, Regenerate the affected MCP servers to pick up the new image.

See spec/deployment-workflow.md for the full workflow, prerequisites, and troubleshooting.

Create a login user (no public registration by design):

./infra/scripts/create-user.sh --username alice

Tech stack

NestJS 11 · React 19 · Mastra + Vercel AI SDK · Prisma + PostgreSQL · Redis + BullMQ · mem0 + OpenSearch · Langfuse · Kubernetes (EKS) · Terraform.

Security

This is a single‑tenant reference implementation. Before any production / public deployment:

  • Rotate all credentials and use a managed secret store (see infra/k8s/02-secrets.yaml.example).
  • Auth is JWT‑based with a global guard; RBAC, multi‑tenant scoping, and OIDC are not yet implemented.
  • Upstream API credentials are stored plaintext in the DB in MVP (encryption‑at‑rest is roadmap).

Report vulnerabilities per SECURITY.md.

Contributing

See CONTRIBUTING.md.

License

Apache License 2.0.

Ver o original no GitHub

Bom para

Relacionados