#Where each AI agent keeps its MCP config
Every agent runtime stores its MCP servers in its own file, under its own key, and a config copied from one into another usually fails without an error. Some runtimes also read a second file inside the repository, for servers that belong to one project. The user-level files and keys below come from Ambit's readers, src/engine/mcp-clients.ts and src/engine/claude-code.ts, so they are the paths Ambit actually opens.
Paths are for macOS and Linux, with ~ as your home directory, which is %USERPROFILE% on Windows. Six runtimes keep their file under %APPDATA% there, and Ambit reads each: Claude Desktop in Claude\claude_desktop_config.json, VS Code in Code\User\mcp.json, Cline and Roo Code in Code\User\globalStorage\<extension>\settings\cline_mcp_settings.json, Zed in Zed\settings.json, and Goose in Block\goose\config\config.yaml.
#User files
| Runtime | File | Where the servers are |
|---|---|---|
| Claude Code | ~/.claude.json |
mcpServers for user scope, and projects.<path>.mcpServers for local scope |
| Claude Desktop | ~/Library/Application Support/Claude/claude_desktop_config.json (macOS), ~/.config/Claude/claude_desktop_config.json (Linux) |
mcpServers |
| Cursor | ~/.cursor/mcp.json |
mcpServers |
| Windsurf | ~/.codeium/windsurf/mcp_config.json |
mcpServers |
| Gemini CLI | ~/.gemini/settings.json |
mcpServers |
| Codex CLI | ~/.codex/config.toml |
[mcp_servers.<name>] tables |
| OpenCode | ~/.config/opencode/opencode.json, or opencode.jsonc |
mcp |
| Cline | VS Code's globalStorage/saoudrizwan.claude-dev/settings/cline_mcp_settings.json |
mcpServers |
| Roo Code | VS Code's globalStorage/rooveterinaryinc.roo-cline/settings/cline_mcp_settings.json |
mcpServers |
| Continue | ~/.continue/config.json |
mcpServers, or experimental.modelContextProtocolServers |
| Zed | ~/.config/zed/settings.json, or ~/Library/Application Support/Zed/settings.json |
context_servers |
| VS Code | ~/Library/Application Support/Code/User/mcp.json (macOS), ~/.config/Code/User/mcp.json (Linux) |
servers |
| Copilot CLI | ~/.copilot/mcp-config.json |
mcpServers |
| Amp | ~/.config/amp/settings.json, or settings.jsonc |
amp.mcpServers, one key with a dot in its name |
| Goose | ~/.config/goose/config.yaml |
extensions, those of type stdio or streamable_http |
| Kiro | ~/.kiro/settings/mcp.json |
mcpServers |
VS Code's globalStorage directory is ~/Library/Application Support/Code/User/globalStorage on macOS and ~/.config/Code/User/globalStorage on Linux.
#Project files
Each is read only when the runtime works in that repository, and is usually checked in with it. Ambit does not read these yet, so a server declared only in one does not reach its graph.
| Runtime | Project file | Where the servers are |
|---|---|---|
| Claude Code | .mcp.json at the repository root |
mcpServers |
| Cursor | .cursor/mcp.json |
mcpServers |
| Gemini CLI | .gemini/settings.json |
mcpServers |
| OpenCode | opencode.json at the project root |
mcp |
| VS Code | .vscode/mcp.json |
servers |
| Amp | .amp/settings.json |
amp.mcpServers |
| Kiro | .kiro/settings/mcp.json |
mcpServers |
#Why the same server breaks when you copy it
Three differences account for most failed copies:
- The key. Most runtimes use
mcpServers. OpenCode usesmcp, Zed usescontext_servers, VS Code usesservers, Amp usesamp.mcpServers, Goose usesextensions, and Codex CLI uses TOML tables namedmcp_servers. A block pasted under the wrong key is valid JSON that the runtime ignores. - The file format. Codex CLI reads TOML and Goose reads YAML, OpenCode accepts JSON with comments, and the rest read plain JSON. A trailing comma or a comment breaks a plain JSON file.
- Where a project's servers live. A server in a project file exists only in that repository, and Claude Code also keeps private per-project servers inside
~/.claude.json, keyed by the project's path. The same server can be on in one repository and absent in another.
#Claude Code
Claude Code has three scopes. User servers, available in every project, are the top-level mcpServers in ~/.claude.json. Local servers, for one project and visible only to you, are in the same file under projects.<path>.mcpServers, keyed by the project's path. Project servers, shared with everyone who clones the repository, are in a .mcp.json at its root.
Ambit reads the user and local servers and counts a server once, wherever it is declared; it does not read .mcp.json yet. It also reads skills, agents and settings under ~/.claude/. CLAUDE_CONFIG and CLAUDE_HOME point Ambit at a different file or directory.
#Claude Desktop
One JSON file with a top-level mcpServers object. The macOS path is under ~/Library/Application Support/Claude/, and the Linux path under ~/.config/Claude/. CLAUDE_DESKTOP_CONFIG overrides it for Ambit.
#Cursor
~/.cursor/mcp.json, with mcpServers at the top, for servers in every project. A project can add its own in .cursor/mcp.json, in the same shape, which Ambit does not read. CURSOR_MCP_CONFIG overrides the global path for Ambit.
#Windsurf
~/.codeium/windsurf/mcp_config.json, with mcpServers at the top. WINDSURF_MCP_CONFIG overrides it for Ambit.
#Gemini CLI
~/.gemini/settings.json. The servers are under mcpServers, beside Gemini CLI's other settings. A project's .gemini/settings.json can declare more, in the same shape, and Ambit does not read it. GEMINI_MCP_CONFIG overrides the global path for Ambit.
#Codex CLI
~/.codex/config.toml, where each server is a table:
[mcp_servers.github]
command = "npx"
args = ["-y", "@modelcontextprotocol/server-github"]
CODEX_MCP_CONFIG overrides it for Ambit.
#OpenCode
~/.config/opencode/opencode.json, or opencode.jsonc when that is the one present. Servers are under mcp, and each can carry enabled: false to stay declared and off. OpenCode is the format Ambit's own config switch writes, and OPENCODE_CONFIG points Ambit at a different file.
OpenCode also reads an opencode.json at a project's root. Ambit does not put its servers on the graph; the Environment section of My Setup reads the project configs of repositories under REPO_PATH and says how far each has drifted from the global one.
#Cline and Roo Code
Both are VS Code extensions and keep their servers in the extension's global storage, in a file named cline_mcp_settings.json with mcpServers at the top. CLINE_MCP_CONFIG and ROO_CODE_MCP_CONFIG override them for Ambit.
#Continue
~/.continue/config.json. Newer versions use mcpServers; older ones nest servers under experimental.modelContextProtocolServers. Ambit reads either. CONTINUE_MCP_CONFIG overrides it.
#Zed
Zed's settings.json, under context_servers. Linux keeps it in ~/.config/zed/, macOS in ~/Library/Application Support/Zed/. ZED_MCP_CONFIG overrides it for Ambit.
#VS Code
VS Code's user-level mcp.json, under servers. Linux keeps it in ~/.config/Code/User/mcp.json, macOS in ~/Library/Application Support/Code/User/mcp.json. A workspace can carry its own in .vscode/mcp.json, in the same shape, which Ambit does not read. VSCODE_MCP_CONFIG overrides the user-level path for Ambit.
#Copilot CLI
GitHub Copilot CLI keeps ~/.copilot/mcp-config.json, with mcpServers at the top. Each entry names its transport in type: local or stdio with a command, or http or sse with a url. copilot mcp disable switches a server off and keeps it configured; Copilot's documentation does not say where that is recorded, so Ambit cannot see it. COPILOT_MCP_CONFIG overrides the path for Ambit.
#Amp
Amp's user settings, ~/.config/amp/settings.json or settings.jsonc. The servers are under "amp.mcpServers", which is one key with a dot in its name, like the settings beside it. Each is a command and args, or a url and headers. A workspace can declare more in .amp/settings.json, which Ambit does not read. AMP_MCP_CONFIG overrides the user-level path for Ambit.
#Goose
~/.config/goose/config.yaml, where each extension is an entry under extensions:
extensions:
filesystem:
type: stdio
enabled: true
cmd: npx
args: ["-y", "@modelcontextprotocol/server-filesystem", "/tmp"]
The command is cmd, and a remote server is type: streamable_http with its address in uri, so an entry copied from an mcpServers file has to be renamed field by field. enabled: false keeps an extension configured and off. Ambit reads the stdio and streamable_http extensions, and the older sse ones; a builtin or platform extension ships inside Goose and is no server anyone added. GOOSE_MCP_CONFIG overrides the path for Ambit.
#Kiro
~/.kiro/settings/mcp.json, with mcpServers at the top: a command, args and env, or a url and headers. disabled: true keeps a server configured and off. A workspace's .kiro/settings/mcp.json can add more, in the same shape, which Ambit does not read. KIRO_MCP_CONFIG overrides the user-level path for Ambit.
#Seeing every runtime's servers at once
Each file answers for one runtime. To see all of them together, with which servers are on, which are failing a check, and what stops working if one goes down:
npx ambit-cli # reads every user-level file above that exists, and says what it found
npx ambit-cli status # which servers are on, what is failing, what has one provider
npx ambit-cli impact mcp:github # what stops, and what only weakens, if this one goes
Nothing leaves the machine, and npm install -g ambit-cli or brew install zz-plant/tap/ambit keeps the ambit command. Or open the hosted demo and paste one of the JSON files into the page, where it is read in the tab and never uploaded.
Related: what breaks if an MCP server goes down and auditing which MCP servers a coding agent has.
See what your own agents can reach
Ambit reads the MCP configs of Claude Code, Cursor, OpenCode and thirteen more agents into one local map: every server, which ones work, what one more step would open, and what breaks if a piece goes.
Open the live demoa sample setup in your browser, nothing to install
On your machine: npx ambit-cli, with Node 22.18 or newer on macOS or Linux.
MIT licensed · one local SQLite file, no telemetry · on npm and in the official MCP registry