agentdrive 0.0.1

Placeholder for AgentDrive's stdio MCP companion. Full release coming soon — see https://agentdrive.mnexa.ai.

agentdrive

v0.0.1 is a name-claim placeholder. Running uvx agentdrive today prints an info message and exits. The full stdio shim (the upload(file_path, ...) tool described below) ships with the next release. Until then, the hosted AgentDrive MCP at https://agentdrive.mnexa.ai/mcp already covers reads, search, and writes that don't need local-disk access.

Stdio MCP shim for AgentDrive. Lets agents on stdio-capable hosts (Claude Desktop, Cursor, Claude Code, Windsurf, Zed) upload local files into your AgentDrive without base64-encoding bytes through the model's token stream.

The hosted AgentDrive MCP at https://agentdrive.mnexa.ai/mcp already provides list_artifacts, read, grep, overview, lookup, and the rest of the tool surface. This shim only adds upload(file_path, ...) — the one tool that benefits from running in the user's process (filesystem access). Use both servers side-by-side.

Install + run

uvx agentdrive

Or:

pipx install agentdrive
agentdrive

Configuration

Environment variables:

Var	Required	Default	Notes
AGENTDRIVE_API_KEY	yes	—	Your ad_live_… key. Get it from agentdrive.mnexa.ai/settings/api-keys.
AGENTDRIVE_BASE_URL	no	https://agentdrive.mnexa.ai	Override for staging/dev.
AGENTDRIVE_TIMEOUT	no	60	Per-upload timeout in seconds.

Host config

Add to your MCP host config alongside the hosted AgentDrive MCP. Example for Claude Code (~/.claude.json global, or .mcp.json per-project — or use claude mcp add):

{
  "mcpServers": {
    "agentdrive": {
      "url": "https://agentdrive.mnexa.ai/mcp",
      "headers": { "Authorization": "Bearer ad_live_..." }
    },
    "agentdrive-upload": {
      "command": "uvx",
      "args": ["agentdrive"],
      "env": { "AGENTDRIVE_API_KEY": "ad_live_..." }
    }
  }
}

The agentdrive server handles reads, search, and listings; the agentdrive-upload server handles file uploads. Your MCP host routes tool calls to whichever server registered each tool name.

Using alongside other MCP servers

MCP hosts spawn every configured server in parallel and merge their tool catalogs. Just add more entries to mcpServers — mix stdio (command + args) and hosted (url + headers) freely:

{
  "mcpServers": {
    "agentdrive":         { "url": "https://agentdrive.mnexa.ai/mcp",
                            "headers": { "Authorization": "Bearer ad_live_..." } },
    "agentdrive-upload":  { "command": "uvx", "args": ["agentdrive"],
                            "env": { "AGENTDRIVE_API_KEY": "ad_live_..." } },

    "filesystem":   { "command": "npx",
                      "args": ["-y", "@modelcontextprotocol/server-filesystem", "~/Documents"] },
    "git":          { "command": "uvx", "args": ["mcp-server-git"] },
    "github":       { "url": "https://api.githubcopilot.com/mcp/",
                      "headers": { "Authorization": "Bearer ghp_..." } },
    "linear":       { "url": "https://mcp.linear.app/sse",
                      "headers": { "Authorization": "Bearer lin_..." } }
  }
}

Three things worth knowing:

• Tool name collisions are namespaced by server. If two servers expose upload, the host disambiguates (e.g. agentdrive-upload__upload vs other__upload). That's also why our two AgentDrive entries are named agentdrive and agentdrive-upload — distinct keys, distinct sources.
• Tool count budget. Each server contributes 5–15 tools; past ~50–80 total in a model's catalog, tool-selection accuracy starts to degrade. Prune what you're not using.
• Config file location varies by host — same JSON shape, different paths:
  • Claude Desktop (macOS): ~/Library/Application Support/Claude/claude_desktop_config.json
  • Claude Code: ~/.claude.json (global, or use claude mcp add) or .mcp.json (per-project)
  • Cursor: ~/.cursor/mcp.json (global) or .cursor/mcp.json (project)
  • Windsurf: ~/.codeium/windsurf/mcp_config.json
  • Zed: project .zed/settings.json under context_servers

The upload tool

upload(
  file_path:       str,                     # local path; required
  path:            str | None = None,       # destination in drive; defaults to filename
  visibility:      "public" | "private" = "public",
  labels:          list[str] | None = None,
  metadata:        dict | None = None,
  source:          dict | None = None,      # v0.6 typed provenance
  actor_name:      str | None = None,
  change_summary:  str | None = None,
  content_type:    str | None = None,       # inferred from extension if omitted
  if_match:        int | None = None,       # optimistic concurrency
) -> dict

Returns the artifact JSON: id, path, url, content_type, size_bytes, version_number, labels, source, etc.

How it works

The shim runs on your machine as a subprocess (stdio MCP server). When an agent calls upload(file_path, ...):

1. The shim reads the bytes from local disk.
2. It PUTs them to https://agentdrive.mnexa.ai/v0/artifacts/{path} over HTTPS with your bearer token in the Authorization header.
3. Hosted AgentDrive stores the artifact, returns the metadata, and the shim returns it to the agent.

Bytes never traverse the model's token stream. The AgentDrive side enforces the usual 50 MB per-artifact cap, rate limits, and quotas.

License

MIT.