What is the recommended practice of working with mutiple projects simultaneously?

[zzJinux]

I typically manage multiple Claude Code sessions across multiple work trees of multiple Git repos. I added the Serena MCP client to the user-level Claude Code config and ended up facing the limitation that one Serena MCP server can have at most one "active project" at a time.

What is the recommended Serena setup for my use case? Is it the only way to run one Serena MCP server process per working tree, or is there an alternative approach?

[zzJinux]

I've added two new components to my custom build: a separate daemon that handles multiple projects and a PROXY stdio mcp server that connects to the daemon.

Usage: I spin up the daemon in advanced (from within a dedicated terminal tab or LaunchAgent) and add the PROXY to Claude Code's user-level mcp server settings. Each session, when it starts, creates a PROXY process that knows which project it's dealing with based on the working directory. I no longer have to worry about which project is currently active.

Asking @opcode81 for PR eligibility. Does this design sound ok to you?

[opcode81]

Can you explain why one Serena process per project being worked on is a practically relevant limitation? What is the underlying problem that you are trying to solve?

[zzJinux]

These are my own requirements

1. Allows Serena-enabled agent sessions that work in parallel in multiple working trees (projects)
2. Easy setup: configure once for user-level MCP setting and forget
3. One language server process per project

With the current Serena build:

• Use the Serena HTTP server for user-level MCP: all agent sessions contact a single Serena process: violates the (1)
• Use the Serena stdio server for user-level MCP: two or more agent sessions will have their own language server process, respectively: violates (3)

[zzJinux]

@ opcode81, is it questionable how it proves the current usage pattern? I thought having multiple language server processes for one working tree is just a waste of resources.

[ketema]

Following... researching threads around this. I use one global serena http server for many agents, but this is limited because only one serena project can be active at a time... I have a different approach than a proxy in mind trying to gauge interest

[xizhibei]

I ran into the same class of problem: stdio gives each agent/session its own Serena process, while one shared HTTP Serena instance is only a clean fit when all clients work on the same active project.

I am working on 1MCP, and this is one reason we added template servers. Instead of hardcoding Serena once per client or copying MCP JSON per repo, you define Serena once as a template:

{
  "mcpTemplates": {
    "serena": {
      "command": "uvx",
      "args": [
        "--from",
        "git+https://github.com/oraios/serena",
        "serena",
        "start-mcp-server",
        "--project",
        "{{project.path}}",
        "--context",
        "claude-code"
      ],
      "tags": ["serena", "semantic", "project"]
    }
  }
}

When a client/session connects, 1MCP renders that template from the current project context. Same project plus same rendered config can reuse the same Serena instance; different projects or worktrees get isolated Serena instances.

This is not a replacement for Serena's own project model. Serena still provides the semantic code tools. 1MCP handles the runtime layer around it: project-aware config, routing, lifecycle, client-specific context, tags, presets, and tool discovery.

Docs:

• 1MCP Serena integration: https://docs.1mcp.app/guide/integrations/serena
• 1MCP repo: https://github.com/1mcp-app/agent

Disclosure: I am involved with 1MCP.