Deploy AI applications

Run MCP servers, AI agents, agent swarms, chatbots, RAG and self-hosted AI tools on Sevalla: how streaming, long tasks, private services and state work, and how to keep them secure.

An AI application runs on Sevalla like any other application: a web process serves its API, chat interface or MCP endpoint, background workers run long agent tasks, and databases keep conversations, queues and embeddings. The models run at providers such as Anthropic and OpenAI, called over HTTPS with API keys in environment variables, or on CPU inside Sevalla with Ollama.

Every guide in this section was followed on Sevalla from an empty folder. Guides for code you write build with Ingot and with a Dockerfile; guides for packaged tools deploy the vendor’s official Docker image. Each one ends with the same deployment as API calls, so an AI agent can run it.

MCP servers

MCP server with TypeScript

A remote MCP server on Streamable HTTP with the official TypeScript SDK, bearer token auth and PostgreSQL tools.

MCP server with Python

The same server with the official Python SDK and uv, stateless so it runs on several instances.

Agents

Claude Agent SDK

An agent with custom tools behind an HTTP API, a background worker and sessions in PostgreSQL.

OpenAI Agents SDK

Agents with tools and a handoff, streamed by FastAPI, with sessions in PostgreSQL.

LangGraph

A tool-calling graph with the PostgreSQL checkpointer and a human approval step.

Mastra

A TypeScript agent with a tool, a workflow and memory in PostgreSQL.

Claude Code

A Claude Code session in the cloud that you open from the web terminal and that keeps working when you leave.

Agent swarm

A builder and a reviewer agent that work through GitHub pull requests, and a queue between them.

Chat and retrieval

AI chatbot

A Next.js chatbot with the AI SDK that streams answers and saves chats in PostgreSQL.

RAG with pgvector

Embeddings in PostgreSQL with an HNSW index, and answers that cite their sources.

Hybrid search

PostgreSQL full-text and vector search merged with reciprocal rank fusion, measured on real queries.

Qdrant

A private vector database on a disk, with snapshots in object storage.

Models and tools you host

Ollama

Open models on CPU in a private application, with measured speeds per size.

LiteLLM

One OpenAI-compatible endpoint for every model, with a key and budget per application.

OpenRouter

Models from many providers through one API, with fallbacks and cost per key.

Open WebUI

A private chat interface for your team, connected to LiteLLM and Ollama.

LibreChat

A chat interface with MongoDB and Meilisearch as private applications, file uploads on a disk and sign-up under your control.

AnythingLLM

Chat with your documents, with storage on a disk and Claude, LiteLLM or Ollama as the model.

n8n

Workflow automation with AI agent nodes, scaled with queue mode and workers.

Langflow

Visual AI flows that you run through Langflow’s API.

Flowise

Visual flows with PostgreSQL and fixed secrets, called through Flowise’s API with an API key.

Dify

Dify from its official images, with a working code sandbox, pgvector and a Claude workflow.

Langfuse

Tracing, cost and evaluation for your model calls, self-hosted.

How AI applications run on Sevalla

Topic	On Sevalla
Model API keys	Store them as environment variables available during runtime only, so they never reach the build or the image.
Streaming	Server-Sent Events, chunked responses and WebSockets reach the client unbuffered through Cloudflare and Sevalla’s proxy, with no extra headers.
Long agent runs	A request that sends nothing for too long gets Cloudflare’s 524 (see HTTP errors). Stream progress so bytes keep flowing, or run the task in a background worker and let the client poll for the result.
Several instances	Sevalla sends each request to any instance, with no session affinity. Keep conversations, sessions and queues in PostgreSQL, Redis or Valkey, and run MCP servers in stateless mode.
Private services	Turn HTTP traffic off for model servers, proxies and agents nobody should reach from the internet, and call them over internal connections.
Vectors	PostgreSQL databases created with Enable pgvector extension store embeddings; the extension can’t be turned on later. See RAG with pgvector.
Files and models	A disk keeps files across deployments, for one instance. Object storage holds files that every instance reads.
Workers	Start a worker’s command with exec, so it gets SIGTERM and finishes its task when an instance stops. See How processes run.
Hardware	CPU only: no process size has a GPU. Small open models run on Ollama; larger ones belong at a model provider. See Limits.
Interactive agents	The web terminal opens a shell in a running instance, and a tmux session keeps an agent working when you close the tab. See Claude Code.

Keep AI applications secure

• Authenticate every public endpoint that can spend money or call tools, at least with a bearer token in an environment variable.
• Never expose an agent with shell or write tools to the internet. Turn its HTTP traffic off and reach it through the web terminal or an internal connection.
• Treat what an agent reads as untrusted. Pull requests, web pages, tickets and tool results can carry prompt injection, so let your code, not the model, take irreversible actions such as merging, paying or deleting.
• Give agents least-privilege tokens, such as a fine-grained GitHub token for one repository, and set spending limits on your model provider accounts.

To let an AI agent deploy and manage these applications for you, connect it to the Sevalla MCP server.