ByteChef: workflows and MCP servers for agents
ByteChef
An open-core platform: workflows and actions of connected services are published as MCP tools for Claude, Cursor and Windsurf
- 1,057 ★ on the repository
- Last commit
Broad access
Why this level
- Server tools perform actions in connected services, such as Slack or HubSpot, using the credentials of your connections
- With authentication off, the secret key in the server address is the only protection
- One tool call runs the whole workflow with all its steps
Install
- Works in
Manual install
docker run --name bytechef -it -p 8080:8080 --env BYTECHEF_DATABASE=h2 --env BYTECHEF_AI_MEMORY_PROVIDER=in_memory -v ~/.bytechef:/root/.bytechef docker.bytechef.io/bytechef/bytechef:latestQuick start for evaluation with the built-in H2 database; sign in at `http://localhost:8080/login`.
Access and checks
How we reviewBroad accessThe tool can change data, run commands or handle secrets.
- Server tools perform actions in connected services, such as Slack or HubSpot, using the credentials of your connections
- With authentication off, the secret key in the server address is the only protection
- One tool call runs the whole workflow with all its steps
This is third-party code. Review the repository files before installing.
What it does
ByteChef is an open-core platform (Apache 2.0) for workflow automation and AI agents that you self-host with Docker. The project claims more than 250 connectors, and each one serves as an agent tool. For Claude Code, Cursor and Windsurf the MCP part matters. In the Automation workspace you create an MCP server and choose what to give the agent: individual component actions (for example, send a Slack message) and whole workflows that start with the New Workflow Call trigger. The server gets an address like `https://<your-host>/api/automation/<secret-key>/mcp`, and the Connect tab provides ready snippets for Claude, Cursor, Windsurf and other clients. Each tool has a name, description and parameters: you fix some yourself and the model fills the rest at call time with the `fromAi` expression. The platform also works as a client: external MCP servers connect to a ByteChef agent as tool sources. Every tool call runs as an ordinary workflow execution and is visible in the execution log with inputs, outputs and steps.
Who it is for. For teams that already automate processes and want to give Claude Code or Cursor ready workflows and service actions without handing the agent your credentials.
Good fit when
- You want the agent to run your process (for example, issuing a refund) as one tool
- The agent needs actions in Slack or HubSpot and you do not want to hand it credentials directly
- You want to see every agent call in the execution log with inputs and outputs
- You need one place for connectors, workflows and agent tools, self-hosted
- You want to change a tool's behavior in one place without editing every client
Not a fit when
- You need one ready MCP server for a specific service: deploying a platform with a database for it is excessive
- You do not want to administer Docker and PostgreSQL: the platform runs on your server
- You need features marked Enterprise Edition, such as admin API keys: they are licensed separately
- The call must finish in seconds: the server waits up to 300 seconds for a workflow to complete
Example request
Call the refund tool from ByteChef for order 4812 and show how the run ended.MCP
- Transport
- http
- Authentication
- API key
Limitations
You have to deploy the platform yourself: a Docker container with the built-in H2 database suits evaluation, and for long-term use the project recommends Docker Compose with PostgreSQL. There is no migration from H2 to PostgreSQL. A new MCP server is created disabled, and only workflows with the New Workflow Call trigger reach the tool list.
Code outside `server/ee/` and `client/src/ee/` is Apache 2.0, while the Enterprise Edition license applies inside them. The server address contains a secret key: with authentication off it is the only protection. Availability of the Docker images from Russia was not checked for this card.
How to disable. Turn off the Enabled switch of the MCP server in the MCP Servers section or delete the server. The refresh button on the Connect tab rotates the secret key in the address, after which the old address stops working. In the client, `claude mcp remove` with your server name removes it.
FAQ
How do I connect ByteChef to Claude Code or Cursor?
Create an MCP server in the Automation section, add the actions and workflows you need, turn it on and copy the address like `https://<your-host>/api/automation/<secret-key>/mcp` from the Connect tab. The same tab has ready snippets for Claude, Cursor and Windsurf.
Is authentication required to connect?
For new servers, yes: an API key or a JWT from a trusted issuer is needed, otherwise the response is 401. Servers created before the setting existed work without authentication until you turn on the Require authentication checkbox.
How do I start ByteChef quickly for evaluation?
You need Docker Desktop. Run `docker run --name bytechef -it -p 8080:8080 --env BYTECHEF_DATABASE=h2 --env BYTECHEF_AI_MEMORY_PROVIDER=in_memory -v ~/.bytechef:/root/.bytechef docker.bytechef.io/bytechef/bytechef:latest` and open `http://localhost:8080/login`. The H2 database is for evaluation only.
Which workflows can be given to an agent?
Only those that start with the New Workflow Call trigger (`workflow/v1/newWorkflowCall`). Each becomes a single tool; the call waits up to 300 seconds and returns the workflow's response. If a workflow stops at an approval step, the tool returns the status `approval_required`.
What is free in ByteChef and what needs the Enterprise Edition?
The core is Apache 2.0, suitable for self-hosting and commercial use. Code in `server/ee/` and `client/src/ee/` is covered by the Enterprise Edition license. That includes, for example, admin API keys.
Discussion
No questions or reviews yet.
Related
MCP serversAn MCP server for driving Windows: the agent opens apps, clicks and types, reads window and UI state, and works with files and PowerShell
apra-fleet runs a fleet of AI agents across machines and models through MCP and drives multi-hour workflows
Zapier's official MCP plugin: the agent gets actions across thousands of apps through your Zapier account
A self-hosted control plane for running several agents: task dispatch, run review, spend tracking, connected via MCP, CLI or REST
Flow-Like MCP
Flow-Like
Flow-Like MCP nodes expose visual workflow functions to an agent over HTTP, with owner-selected tools and access
MCP server and wincli CLI: agents drive Windows apps by control name through UI Automation instead of screen coordinates
Comparison with similar tools
Cards that share the main search query, and how they differ.
| Tool | Official | Authentication | Repository stars | Main difference |
|---|---|---|---|---|
| ByteChef: workflows and MCP servers for agents (this card) | yes | API key | 1,057 | |
| n8n-MCP | no | key is optional | 23,220 | an n8n node reference for the agent; ByteChef itself hosts workflows and exposes them as MCP tools |
| Zapier MCP | yes | OAuth | 453 | access to app actions through the Zapier cloud; ByteChef is self-hosted |
README in short
ByteChef is an Apache 2.0 alternative to n8n and Zapier. Differences: every run is persisted and resumes where it stopped, an embedded mode for SaaS products, and an agent as a step with a loop. It continues with a feature table by edition, 12 model providers, memory and vector stores, agent guardrails, Docker launch options and answers to common questions.