How to deploy an app from the Zed agent in 2026

Zed's agent deploys an app once you give it a hosting provider's MCP server. Add the server under context_servers in Zed's settings file, finish the browser sign-in Zed starts for remote servers, then ask the Agent Panel to deploy the open project. Pick Eurobase when the app has to run on European infrastructure, Vercel for Next.js on its global network, and Render or Railway when you want the agent to set up a database alongside the app.

Hosting servers you can add to Zed at a glance

Platform	Server address	How Zed signs in	What the agent can do	Where apps run
Eurobase	https://api.eurobase.dev/mcp	OAuth in your browser	Create a project from your repository, deploy, roll back, publish Functions	European infrastructure
Vercel	https://mcp.vercel.com	OAuth	Manage projects and deployments, read deployment logs	Global
Render	https://mcp.render.com/mcp	API key in a header	Create web services, static sites, Postgres and Key Value, trigger deploys	Oregon, Ohio, Virginia, Frankfurt, Singapore
Railway	https://mcp.railway.com	OAuth, or the Railway CLI	Create projects, deploy templates, redeploy services	California, Virginia, Amsterdam, Singapore

Server details are from each provider's own MCP page, checked on 10 November 2026.

How Zed connects to an MCP server

Zed calls MCP servers "context servers". Open Settings, then AI, then MCP Servers, click Add Server and choose Add Remote Server. Zed writes the entry to its settings file, which you can also edit by hand with the zed: open settings file action. A hosted server needs only a URL:

"context_servers": { "eurobase": { "url": "https://api.eurobase.dev/mcp" } }

If the entry has no Authorization header, Zed starts the standard MCP OAuth flow and asks you to sign in. A server that wants an API key instead takes it under headers, for example "Authorization": "Bearer <token>".

Back on the MCP Servers page, a green dot with the tooltip "Server is active" means the connection works. Zed supports the Tools and Prompts parts of MCP, and it reloads a server's tool list on its own when the server changes it.

Keep deploys behind an approval

Since Zed v0.224.0, agent.tool_permissions.default decides what happens when the agent wants to run a tool. The default is confirm, so Zed asks before every action, MCP calls included. You can switch to allow or deny, and you can set a rule for one tool with the key mcp:<server>:<tool_name>.

A sensible setup for hosting servers is to leave reads on allow and keep anything that deploys, promotes or rolls back on confirm. If the agent keeps reaching for the terminal instead of the hosting tools, mention the server by name in your prompt, or build an Agent Profile that turns off the built-in tools you don't want and enables only that server's tools.

Zed's own agent or an external one

Zed runs two kinds of agent in the same panel. The built-in Zed Agent uses the context servers you configured directly. External agents such as Claude Agent, Codex, OpenCode or Cursor connect through the Agent Client Protocol. Zed can pass your configured MCP servers to them over that protocol, and they can also load servers from their own MCP config. Zed Agent profiles don't apply to them, and they handle their own sign-in and billing.

If you already connected a host in Claude Code or Codex, the external agent can use that connection from its own config. For those setups, see deploying from Claude Code and deploying from Codex.

What to check before the agent ships anything

• Token or browser sign-in. OAuth keeps keys out of settings.json. A header token sits in the file in plain text.
• How far access reaches. Some servers act as your whole account, others let you pick projects.
• Where secrets end up. You don't want database passwords typed into a chat with a model.
• Whether the agent sees build output. A failed deploy is only fixable if the server returns logs.
• The region. The agent deploys to wherever the platform runs, so the hosting question doesn't change.

1. Eurobase

Eurobase hosts web apps, Workers and Functions on European infrastructure run by Scaleway, a French provider, and the business is registered in Warsaw, Poland. Its docs give the exact Zed entry shown above: add it to settings.json, then sign in when Zed asks.

The sign-in page opens in your browser. You choose the projects the agent may use, the workspaces where it may create new ones, and read or write for each permission. Then ask: "Deploy this project on Eurobase." The agent creates the project from your repository, starts the first deployment and follows it until the site answers at a .eurobase.dev address. With write access it can also redeploy, promote, roll back, set public variables and publish Functions. Deleting is outside its reach. Secrets stay with you too: you enter their values in the project's Secrets, and the agent never sees them.

Deleting the context_servers entry leaves the grant in place. To end access, go to Connected apps; the agent is cut off at its next request.

• Best for: apps from Zed that should stay on European infrastructure, with the agent limited to the projects you pick.
• Price: a Free plan, then €10 or €20 a month + VAT. There are no usage charges today: work over a limit is refused, apps on V8 isolates and Functions stop until the monthly reset, and containers and static sites keep serving.
• What it lacks: preview deployments, managed Postgres, MySQL or Redis, and Dockerfile builds. Zed uses the same sign-in as Claude Code and Codex, but only those two are tested end to end so far. ISO 27001 and SOC 2 certification is still underway.

2. Vercel

Vercel's official server is a remote MCP with OAuth at https://mcp.vercel.com, available on all Vercel plans. Zed isn't in Vercel's list of supported clients, but Vercel says its npx add-mcp setup command also configures other HTTP clients, Zed among them. You can also paste the URL under context_servers and sign in.

• Best for: Next.js projects you want on Vercel.
• What it lacks: a European company, since Vercel Inc. works under California law. Vercel also advises keeping human confirmation on, which in Zed means leaving its tools on confirm.

3. Render

Render links to Zed's MCP docs from its own setup page and points out that Zed uses context_servers where other tools use mcpServers. Its documented examples for generic clients pass a Render API key as a bearer header. The server can create web services, static sites, cron jobs, Render Postgres and Render Key Value, trigger deploys, update environment variables and read logs and metrics.

• Best for: an app and its Postgres database set up from one conversation.
• What it lacks: image-backed services and IP allowlists, which Render says the MCP server can't create. It also warns it can't guarantee that sensitive values stay out of the agent's context.

4. Railway

Railway's server at mcp.railway.com creates projects, deploys templates, manages environments, pulls variables and redeploys services. Editors that support OAuth can connect to it directly, and Zed does, though Railway's per-editor instructions don't include a Zed entry. The other route runs railway mcp through the Railway CLI, version 5.44.0 or later, which reuses your railway login.

• Best for: agents assembling a stack from Railway templates.
• What it lacks: Zed-specific setup steps in Railway's own docs.

A first deploy from Zed

1. Push the project to a Git repository the host can read.
2. Add one hosting server under context_servers and sign in. One server keeps the agent from picking the wrong platform.
3. Check the green dot on the MCP Servers page.
4. Open the Agent Panel and ask it to deploy and send back the live URL.
5. Approve the deploy call, then add any secrets in the host's dashboard and ask for a redeploy.

Frequently asked questions

Where is Zed's settings file for MCP servers?

Run zed: open settings file from the command palette. On macOS, and on Linux without a custom config directory, it's ~/.config/zed/settings.json. Servers go under the context_servers key, not mcpServers.

Why isn't Zed's agent using my hosting server?

Check the dot next to the server on the MCP Servers page first. If it's green, name the server in your prompt. Zed notes that how reliably MCP tools get called varies by model, and a profile that enables only that server's tools removes the choice.

Can I deploy from Zed with Claude or Codex instead of the Zed Agent?

Yes. Install them from the ACP Registry as external agents. Zed can pass its MCP servers to them, and they can also use servers set up in their own config, such as one you added with claude mcp add or codex mcp add.

Does Zed ask before the agent deploys?

By default, yes. agent.tool_permissions.default is confirm unless you change it, and that covers MCP tool calls. You can loosen it per tool with mcp:<server>:<tool_name>.

How do I deploy from Zed to a European host?

Connect a host whose infrastructure is in Europe. Eurobase runs only there. Render and Railway each offer one European region you can choose. For a wider comparison, read cloud MCP servers compared.

Sources

• Zed: Model Context Protocol, Tool Permissions, Agent Panel and External Agents
• Vercel MCP
• Render MCP server
• Railway MCP server
• Eurobase AI agents, Deploy with Claude Code, Codex or Cursor, Eurobase pricing and deploying from Cursor