Setup by IDE

Feri works with any IDE or agent that supports MCP servers and runs a tool-calling session. Here is exactly where the MCP server, skill, and steering go on each one.

How the pieces map

Feri has three parts. Only the first is required to run:

• The MCP server (required): the Feri process your IDE launches and calls as a tool. Every supported IDE registers this the same way, in its own MCP config file.
• The skill (skill/skill.md): documents the tools and how to use them.
• The steering file (steering/feri.md): sets runtime behaviour, like polling cadence and stop handling.

Skills and steering are native, separate concepts in Kiro. Other IDEs have one mechanism for persistent custom instructions (rules files, custom instructions, guidelines). On those, put the contents of both the skill and the steering file into that mechanism and set it to always apply. The behaviour is the same; only the home differs.

Compatibility

The requirement is MCP support plus an agent that can call tools between tasks. Tools without MCP are not supported (see the bottom of this page).

1. Add the MCP server

First build Feri (npm install && npm run build), then register the built entry point as a stdio MCP server. The standard shape, used by most IDEs:

{
  "mcpServers": {
    "feri": {
      "command": "node",
      "args": ["/absolute/path/to/feri/dist/index.js"]
    }
  }
}

No secrets go in this file. Feri reads its own config and connection code at boot. Use an absolute path to dist/index.js.

The root key is not the same everywhere. Get it wrong and the server silently won't load:

• Most IDEs use mcpServers (the shape above).
• VS Code + Copilot and Visual Studio use servers.
• Zed uses context_servers (in settings.json).
• Continue uses YAML, not JSON (see below).
• Goose registers MCP as an extensions entry in config.yaml.

VS Code / Visual Studio (servers key):

{
  "servers": {
    "feri": { "command": "node", "args": ["/absolute/path/to/feri/dist/index.js"] }
  }
}

Zed (context_servers in settings.json):

{
  "context_servers": {
    "feri": {
      "command": { "path": "node", "args": ["/absolute/path/to/feri/dist/index.js"] }
    }
  }
}

Continue (.continue/mcpServers/feri.yaml):

name: Feri
version: 0.0.1
schema: v1
mcpServers:
  - name: feri
    command: node
    args:
      - /absolute/path/to/feri/dist/index.js

Claude Code (CLI, instead of editing a file):

claude mcp add feri -- node /absolute/path/to/feri/dist/index.js

2. Add the skill + steering

On Kiro, drop skill/skill.md into your skills directory and steering/feri.md into .kiro/steering/. They load as a first-class skill and steering file.

On every other IDE, paste the contents of both files into that IDE's persistent rules / instructions mechanism and set it to always apply. Where each one lives:

Per-IDE MCP reference

Exact MCP config file and root key for each supported IDE:

Workspace-scoped config beats user/global config where both exist. Visual Studio also auto-discovers .vscode/mcp.json and .cursor/mcp.json, so a repo set up for another IDE may work without changes.

Not supported

Feri needs an MCP client and an agent tool-loop. Tools that only do inline completion or chat, with no MCP client, can't run Feri:

• GitHub Copilot classic completions (inline autocomplete, non-agent). MCP only exists in Copilot agent mode inside VS Code, Visual Studio, and JetBrains, which are supported.
• Tabnine: no MCP client in its documentation.
• Legacy Codeium plugin (the pre-Windsurf VS Code / JetBrains plugin). MCP is in the Windsurf editor, which is supported.
• Replit Agent: Replit's docs cover building and hosting MCP servers, not the Agent consuming external stdio MCP servers in-session. Treat as unsupported until confirmed.

If your IDE has MCP support that isn't listed here, it most likely works. Open an issue on GitHub and we'll confirm it.