Use obsei from your AI tool

Every tool runs the same local MCP server over stdio:

uvx --from "obsei>=1.0" obsei mcp

uvx comes with uv. The >=1.0 floor stops uvx from reusing an older cached obsei, such as a release candidate. The first start downloads obsei, so give the server a startup timeout of about 60 seconds where the tool has one.

The server reads obsei.yaml in the working directory, or the file named by OBSEI_CONFIG. Set OBSEI_DB for a database somewhere else and OBSEI_DB_KEY when it is encrypted. It opens the store read-only. The plugins for Claude Code, Codex, Copilot, VS Code, Cursor and Gemini also add the voice-of-customer skill: one SKILL.md shared by all of them.

Tool	Install	Notes
Claude Code	/plugin install obsei --marketplace obsei/obsei	MCP server and skill. Older versions: /plugin marketplace add obsei/obsei, then /plugin install obsei@obsei
Codex	codex plugin marketplace add obsei/obsei, then codex plugin add obsei@obsei	MCP server and skill. MCP only: below
GitHub Copilot CLI	copilot plugin install obsei/obsei:integrations/claude/obsei	MCP server and skill
VS Code (Copilot agent mode)	button above, or code --add-mcp '{"name":"obsei","command":"uvx","args":["--from","obsei>=1.0","obsei","mcp"]}'	MCP server. For the skill too, run Chat: Install Plugin From Source with https://github.com/obsei/obsei
Cursor	button above	MCP server. The plugin with the skill is awaiting Cursor Marketplace review
Gemini CLI	gemini extensions install https://github.com/obsei/obsei	Asks for OBSEI_CONFIG, OBSEI_DB and OBSEI_DB_KEY; Gemini passes only these to the server
Antigravity	agy plugin import gemini	Imports the Gemini CLI extension above
Claude Desktop	JSON config in claude_desktop_config.json
Cline	ask Cline to install https://github.com/obsei/obsei	It follows llms-install.md
Devin Desktop (formerly Windsurf)	JSON config in ~/.config/devin/mcp_config.json
Zed	Zed settings
JetBrains AI Assistant	Settings > Tools > AI Assistant > Model Context Protocol (MCP), Add, paste the JSON config
Junie	JSON config in .junie/mcp/mcp.json
Kiro	JSON config in .kiro/settings/mcp.json
Goose	Install in Goose	Or goose configure, Add Extension, Command-line Extension with the command above
ChatGPT	Secure MCP Tunnel	Private use only

JSON config

Most tools take the same mcpServers block:

{
  "mcpServers": {
    "obsei": {
      "command": "uvx",
      "args": ["--from", "obsei>=1.0", "obsei", "mcp"],
      "env": { "OBSEI_CONFIG": "/path/to/obsei.yaml", "OBSEI_DB_KEY": "..." }
    }
  }
}

VS Code’s .vscode/mcp.json uses servers in place of mcpServers.

Codex

codex mcp add obsei --env OBSEI_CONFIG=/path/to/obsei.yaml -- uvx --from "obsei>=1.0" obsei mcp

Then raise the startup timeout in ~/.codex/config.toml, and forward the database key from your environment rather than writing it to the file:

[mcp_servers.obsei]
command = "uvx"
args = ["--from", "obsei>=1.0", "obsei", "mcp"]
env = { OBSEI_CONFIG = "/path/to/obsei.yaml" }
env_vars = ["OBSEI_DB_KEY"]
startup_timeout_sec = 60

Zed

In settings.json:

{
  "context_servers": {
    "obsei": {
      "command": "uvx",
      "args": ["--from", "obsei>=1.0", "obsei", "mcp"],
      "env": { "OBSEI_CONFIG": "/path/to/obsei.yaml" }
    }
  }
}

ChatGPT

ChatGPT does not run local stdio servers. Its connectors need a public HTTPS MCP server with OAuth 2.1 (or no authentication) and reject static bearer tokens, so obsei serve with OBSEI_API_TOKEN cannot be added. obsei does not provide an OAuth server yet.

For private use in your own workspace, OpenAI’s Secure MCP Tunnel wraps obsei mcp with an outbound-only connection. Create a tunnel in Platform tunnel settings, download tunnel-client, then:

export CONTROL_PLANE_API_KEY="sk-..."
tunnel-client init --sample sample_mcp_stdio_local --profile obsei \
  --tunnel-id tunnel_... --mcp-command 'uvx --from "obsei>=1.0" obsei mcp'
tunnel-client run --profile obsei

In ChatGPT, open Plugins, select +, Add custom MCP server, choose Tunnel under Connection and pick the tunnel. Keep tunnel-client run running while you use it. Records leave your machine for OpenAI as each tool returns them, already redacted.

Remote MCP

For tools that connect over HTTP with a header, obsei serve serves the same tools at /mcp; see MCP and agents.