AgentDrive

What AgentDrive does

AgentDrive provides a shared, portable workspace for agents operating across different tools and sessions. It stores Markdown artifacts with automatic versioning, supports semantic search and persistent facts, and records which agent platform contributed to a run. Session tools expose recent work and compact handoffs so another harness can continue from the prior goal, touched files, and recent turns.

The workspace also supports human collaboration. Agents can create share links, add comments, read feedback, invite organization members, and manage per-workspace roles. Organization tools expose information about agents, members, plan, and usage.

How it works

The server can run locally over stdio with the @adrive/mcp-server npm package or remotely through https://mcp.agentdrive.sh/mcp using Streamable HTTP and bearer-token authentication. A local instance identifies the active project from its working directory, using the Git remote when available and otherwise the repository path. That project identity scopes session listing and resumption.

Agent platform and role values can be supplied to distinguish activity from different runtimes or agent roles. Supported platform labels include Claude Code, Codex, OpenCode, Gemini CLI, Cursor, Antigravity, LangGraph, CrewAI, AutoGen, Hermes, OpenClaw, and unknown.

Setup and configuration

Create an API key through the AgentDrive application, then configure either the hosted endpoint with an Authorization: Bearer header or the local command below:

{
  "command": "npx",
  "args": ["-y", "@adrive/mcp-server"],
  "env": {
    "AGENTDRIVE_API_KEY": "agentdrive_live_xxx"
  }
}

The key may also be stored through the AgentDrive CLI configuration at ~/.config/agentdrive/config.json. Without a key, the server can start with a browser-based connection tool; the complete toolset becomes available after authentication. AGENTDRIVE_API_URL can point to a self-hosted worker, while AGENTDRIVE_AGENT_PLATFORM and AGENTDRIVE_AGENT_ROLE add provenance metadata.

The repository also provides a Claude Code plugin that captures sessions and supports login through /agentdrive:login. A shell installer can configure detected hosts automatically.

Tools and capabilities

• Retrieve relevant prior artifacts, search organization content semantically, and recall saved facts or preferences.
• List project sessions and resume a handoff from another harness.
• Create, read, list, delete, version, share, comment on, and review files.
• Inspect recent workspace activity and log agent runs with task, status, artifacts, and identity data.
• Invite organization members and manage workspace roles.
• Clone a Git repository into a workspace-specific Linux sandbox, execute shell commands, read sandbox files, and commit changes.

Limitations and notes

The drive_sandbox_* tools are experimental and require the Cloudflare Containers plan on the AgentDrive worker. Without that configuration, they return a not-configured response rather than providing the sandbox workflow. The free tier includes one workspace, 1 GB of storage, and 10,000 embeddings.