# Connect your agent

`logMcp` is log's MCP server. It gives your agent every `log` command as a tool. The installer puts it on your PATH next to `log`.

## Claude Code

```bash
claude mcp add log -- logMcp
```

You add it once. `logMcp` reloads its own code whenever log changes, so you never restart Claude Code for new code (rule 167):
- It waits until no call or job is running, swaps the server underneath, and tells Claude Code the tools changed.
- A tool added since the session started is always reachable through `logCall`.
- `LOG_MCP_RELOAD=off` turns this off.

## Codex

```bash
codex mcp add log -- logMcp
```

## Cursor

In `~/.cursor/mcp.json` (or `.cursor/mcp.json` in a project):

```json
{
  "mcpServers": {
    "log": { "command": "logMcp" }
  }
}
```

## Antigravity (App, CLI, IDE)

In `.agents/mcp_config.json` (workspace-level) or `~/.gemini/config/mcp_config.json` (global):

```json
{
  "mcpServers": {
    "log": {
      "command": "bun",
      "args": ["bin/logMcp.ts"],
      "timeoutSeconds": 300
    }
  }
}
```

Or if `logMcp` is installed on your `PATH`:

```json
{
  "mcpServers": {
    "log": {
      "command": "logMcp"
    }
  }
}
```

In Antigravity IDE or App, you can also open the agent panel's top-right menu (**...**) > **MCP Servers** > **Manage MCP Servers** > **View raw config** to edit `mcp_config.json` directly.

## Any other MCP client

Run `logMcp` as a stdio server, with no arguments. It reads your settings from `~/.logCli/.env`, like `log` does.

## Over HTTP

For a client that can't start a process, serve the tools over HTTP:

```bash
logMcp --http --port 7337
claude mcp add --transport http log http://127.0.0.1:7337/mcp
```

`--host` and `--port` change where it listens. Off your own machine (not `127.0.0.1`) it won't start without a token: set `LOG_MCP_TOKEN` in `~/.logCli/.env`, and clients send `Authorization: Bearer <token>`.

## First call

Your agent's first call is `getGuide({ topic: "logMcpInstructions" })`. It lists every use case and how to ask for it. `getGuide({ useCase: "<platform>/<page>" })` reads one [use case](useCases/README.md). Every tool's description names its guide, the use cases that call it and the AGENTS.md rules it falls under; AGENTS.md itself is the resource `log://agents` ([MCP server](specs/mcpServer.md)).

## Find and run any command

Three tools reach every command, so your agent doesn't have to know which tool to call:

- `logSearch { "query": "send a dm on x" }` finds the commands for a job, in plain words.
- `logGet { "id": "twitter.dm" }` gives the command's inputs and a call that works.
- `logCall { "id": "twitter.dm", "inputs": { … } }` runs it, through the same checks as the command's own tool.

A command that acts on a platform needs `live: true` and `confirmLive: true`. More: [the catalog guide](guides/logCatalog.md).

## Watch a long call

Calls that take a while write one line per step to `~/.logCli/mcp/<day>.log`. Follow along with:

```bash
tail -f ~/.logCli/mcp/$(date -u +%F).log
```

Next: [using log well](usingLog.md).