MCP Configuration

Model Context Protocol (MCP) connects your AI agent to external tools, APIs, and services. The Desktop App (Code Mode), VS Code Extension, and CLI all share a single MCP configuration file, so you only set things up once.

Config File Locations

Level	Path	Scope
User	`~/.abacusai/mcp.json`	All projects for the current user
Project	`./.abacusai/mcp.json`	Current project directory only

Servers from both files are merged. When the same server name appears in both, the user-level entry is used.

Format

Each entry under mcpServers is a named server. There are two types:

Local (stdio) - runs a command on your machine. Uses command, args, and optionally env.
Remote (HTTP) - connects to a network endpoint. Uses url and optionally headers.

{
  "mcpServers": {
    "github": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-github"],
      "env": {
        "GITHUB_PERSONAL_ACCESS_TOKEN": "<YOUR_TOKEN>"
      }
    },
    "brave-search": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-brave-search"],
      "env": {
        "BRAVE_API_KEY": "<YOUR_KEY>"
      }
    },
    "analytics-api": {
      "url": "https://api.example.com/mcp",
      "headers": {
        "Authorization": "Bearer <YOUR_TOKEN>"
      }
    }
  }
}

Field Reference

Local servers (stdio)

Field	Type	Required	Description
`command`	string	Yes	Executable to run (`npx`, `python`, `node`, etc.)
`args`	string[]	Yes	Arguments passed to the command
`env`	object	No	Environment variables (API keys, tokens)

Remote servers (HTTP)

Field	Type	Required	Description
`url`	string	Yes	URL of the remote MCP server
`headers`	object	No	HTTP headers for authentication

Other Ways to Add Servers

Method	Where
`/mcp add`, `/mcp remove`, `/mcp list`, `/mcp auth`	Inside a CLI session. Writes to `mcp.json`.
`abacusai --mcp-config <file>`	Adds a config file for one session
MCP Servers section of the Abacus Settings page in VS Code	Add, edit and remove servers, or open `mcp.json`
"External MCP configs" setting	Also read servers you already set up in Cursor, VS Code, Claude Desktop or Windsurf. Off by default. See Settings.

To allow or block a specific MCP tool without being asked, add a permission rule such as mcp__github__create_issue.

Finding MCP Servers

Browse available servers at:

Best Practices

Store tokens in env or headers rather than hardcoding them in args.
Start with one or two servers. Too many tools can overwhelm the model context.
For team projects, commit .abacusai/mcp.json with project-specific servers but leave secrets out. Each developer supplies their own keys in the user-level file.
Verify a server works by asking the agent a simple question that uses it.

Troubleshooting

If a server isn't connecting:

Check that the command or url is correct.
Make sure required API keys/tokens are filled in (env or headers).
For local servers, confirm the command is in your PATH (npx, node, python).
Validate the JSON. A missing comma or bracket is the most common issue.

For further help, reach out at support@abacus.ai.

Config File Locations​

Format​

Field Reference​

Other Ways to Add Servers​

Finding MCP Servers​

Best Practices​

Troubleshooting​