AgentDrive MCP Server
# AgentDrive MCP Server
The official [Model Context Protocol](https://modelcontextprotocol.io) server for
**[AgentDrive](https://agentdrive.sh)**: shared workspace for coding agents.
Capture sessions from Claude Code, Codex, OpenCode, Cursor, Gemini CLI and
Antigravity, resume them in any of the others, and search everything they wrote.
Provenance on every artifact; portable storage on R2, S3 or your own bucket.
Hit a usage limit in one agent? Call `drive_resume` in another and it loads the
handoff: the goal, the files touched and the latest turns. 22 MCP tools (18
stable + 4 experimental sandbox), over stdio or hosted. Free plan, no credit card.
- **Website:** https://agentdrive.sh
- **Docs:** https://agentdrive.sh/docs
- **npm:** [`@adrive/mcp-server`](https://www.npmjs.com/package/@adrive/mcp-server)
- **Hosted endpoint:** `https://mcp.agentdrive.sh/mcp`
## Claude Code plugin
This repo is also the Claude Code plugin marketplace. The plugin captures every
session in the background and resumes it from Codex, OpenCode, Gemini CLI, Antigravity, or Cursor:
```
/plugin install agentdrive --marketplace agentdrive-sh/agentdrive-mcp
```
On Claude Code older than 2.1.275, run `/plugin marketplace add agentdrive-sh/agentdrive-mcp`
then `/plugin install agentdrive@agentdrive`. Sign in with `/agentdrive:login`.
## Get an API key (free)
Create one at **https://app.agentdrive.sh** — no credit card. The free tier
(1 workspace, 1 GB storage, 10k embeddings) is free indefinitely.
## Two ways to connect
### 1. Hosted (no install) — Streamable HTTP
Point any MCP client that supports remote servers at the hosted endpoint and pass
your key as a bearer token:
```
URL: https://mcp.agentdrive.sh/mcp
Header: Authorization: Bearer agentdrive_live_xxx
```
Or install in one click via **[Smithery](https://smithery.ai/servers/rakesh1002/agentdrive)**.
### 2. Local (stdio) — via npx
```jsonc
{
"mcpServers": {
"agentdrive": {
"command": "npx",
"args": ["-y", "@adrive/mcp-server"],
"env": {
"AGENTDRIVE_API_KEY": "agentdrive_live_xxx"
}
}
}
}
```
Add that block to your host's MCP config:
- **Claude Desktop:** `claude_desktop_config.json`
- **Cursor:** `~/.cursor/mcp.json`
- **Claude Code:** install the plugin (see above), which also captures sessions
- **Codex / Cline / Continue / Zed:** the standard `mcpServers` block above
The fastest path is the AgentDrive installer, which auto-wires every detected host:
```sh
curl -fsSL agentdrive.sh/install | sh
```
## Configuration
| Env var | Required | Default | Description |
|---|---|---|---|
| `AGENTDRIVE_API_KEY` | no | key saved by `adrive auth login` | Your AgentDrive API key (`agentdrive_live_...`). Falls back to `~/.config/agentdrive/config.json`, so client configs never need the key inline. With no key at all the server still starts and offers one tool, `drive_connect`, which signs the user in through the browser and connects their other agent tools; the full toolset then appears in the same session. |
| `AGENTDRIVE_API_URL` | no | `https://app.agentdrive.sh` | Override the API endpoint (self-hosted worker). |
| `AGENTDRIVE_AGENT_PLATFORM` | no | `unknown` | Which agent runtime is calling — drives cross-agent provenance. One of `claude-code`, `codex`, `opencode`, `gemini-cli`, `cursor`, `antigravity`, `langgraph`, `crewai`, `autogen`, `hermes`, `openclaw`, `unknown`. |
| `AGENTDRIVE_AGENT_ROLE` | no | — | Free-text role label (e.g. `planner`, `reviewer`) to distinguish multiple agents on the same platform. |
## Tools
22 tools across files, memory, sessions, collaboration, org, and provenance. The stdio server derives the current project from its working directory (git remote, else repo path), which scopes `drive_sessions` and `drive_resume`.
| Tool | Category | Description |
|---|---|---|
| `drive_context` | memory | Auto-pull the most relevant prior artifacts for the task you're starting. |
| `drive_search` | memory | Semantic search across your org's artifacts. |
| `drive_recall` | memory | Recall persistent facts/preferences agents have written. |
| `drive_sessions` | sessions | List recent agent sessions for this project across Claude Code, Codex, OpenCode, Gemini CLI, Antigravity, and Cursor. |
| `drive_resume` | sessions | Load a compact handoff (goal, files touched, recent turns) from a prior session, by default the latest from a different harness. |
| `drive_write` | files | Write a markdown file (auto-versioned). |
| `drive_read` | files | Read a file's content. |
| `drive_list` | files | List files in a workspace. |
| `drive_delete` | files | Delete a file. |
| `drive_versions` | files | View a file's version history. |
| `drive_share` | collaboration | Generate a (optionally email-gated) share link. |
| `drive_comment` | collaboration | Comment on a file. |
| `drive_feedback` | collaboration | Read human comments on a file. |
| `drive_activity` | provenance | Recent workspace activity (writes, comments, shares). |
| `drive_log_run` | provenance | Log an agent run (task, status, artifacts, identity). |
| `drive_invite` | org | Invite a human to your org by email. |
| `drive_roles` | org | Assign/list/remove per-workspace access by email. |
| `drive_org` | org | View org details: agents, members, plan, usage. |
| `drive_sandbox_clone` | sandbox (experimental) | Clone a git repo into a per-workspace Linux sandbox. |
| `drive_sandbox_exec` | sandbox (experimental) | Run a shell command in the sandbox. |
| `drive_sandbox_read` | sandbox (experimental) | Read a file from the sandbox filesystem. |
| `drive_sandbox_commit` | sandbox (experimental) | Stage + commit the sandbox worktree. |
The `drive_sandbox_*` tools require the Cloudflare Containers plan on the AgentDrive
worker and degrade cleanly to a "not configured" message otherwise.
## Why AgentDrive
Agents lose context at every session boundary and every tool switch. AgentDrive is
the shared substrate: one workspace, every agent, full provenance. Write once with
Claude Code; recall it from Cursor; hand off to Codex — each call is attributed to
the right principal and platform.
## Develop
```sh
npm install
npm run build # tsc -> dist/
npm test # vitest
```
## License
Apache-2.0. See [LICENSE](./LICENSE).
TDQS
Scored across 20 tools
Tools are generally distinct, with clear verb-noun names. Some overlap exists between drive_search, drive_recall, and drive_context, as all retrieve information; however, descriptions clarify their intended use cases.
All tools follow a consistent 'drive_' prefix with snake_case verb-noun or noun-verb pattern. No mixed conventions or irregular naming.
20 tools cover a broad range of workspace management, collaboration, and experimental sandbox features. The count is slightly high but justified by the diverse functionality.
Core file operations (CRUD, sharing, commenting, versioning) are covered, along with user management, search, and context retrieval. Missing explicit workspace creation/renaming, but write and list imply workspace existence.