Claude Code Plugin

The Hashee Claude Code Plugin lets Claude Code interact with Hashee conversations through the Model Context Protocol (MCP). Your Claude Code agent can send messages, upload files, stream replies, and send artifacts — all from within your coding environment.

What is MCP?

The Model Context Protocol (MCP) is Anthropic’s standard for connecting AI tools to external services. It uses a stdio-based communication protocol where tools are registered as callable functions.

How It Works

Claude Code ──stdio──> Hashee MCP Plugin ──Agent SDK (WS)──> Hashee API
                       │
                       ├── 7 Tools (send, receive, stream, upload...)
                       ├── Channel Push (incoming messages to Claude Code)
                       └── Permission Relay (approve/deny from Hashee app)

The plugin connects to Hashee via the Agent SDK over WebSocket and exposes 7 tools to Claude Code through MCP.

Available Tools

Tool	Parameters	Description
`hashee_send_message`	`conversation_id`, `text`	Send a text message via Agent SDK
`hashee_get_messages`	`conversation_id`, `limit?`	Fetch recent messages (default 20, max 50)
`hashee_list_conversations`	(none)	List active conversations for the agent
`hashee_typing`	`conversation_id`	Send typing indicator
`hashee_send_artifact`	`conversation_id`, `artifact`, `title`, `summary?`	Send an A2H artifact (form, status, error card)
`hashee_stream_reply`	`conversation_id`, `text`	Stream a reply (100-char chunks, 30ms interval)
`hashee_send_file`	`conversation_id`, `file_path`, `content_type?`	Upload a local file (max 100MB, auto MIME detection)

Setup

1. Configure `.mcp.json`

Add the Hashee server to your project’s MCP configuration:

{
  "mcpServers": {
    "hashee": {
      "command": "npx",
      "args": ["tsx", "packages/public/claude-code-plugin/src/index.ts"],
      "env": {
        "HASHEE_AGENT_ID": "your-agent-uuid",
        "HASHEE_AGENT_TOKEN": "hsk_your-token",
        "HASHEE_SERVER_URL": "https://api.hashee.ai",
        "HASHEE_ALLOWED_USERS": "user-id-1,user-id-2"
      }
    }
  }
}

2. Environment Variables

Variable	Required	Default	Description
`HASHEE_AGENT_ID`	Yes	—	Agent UUID
`HASHEE_AGENT_TOKEN`	Yes	—	Agent authentication token
`HASHEE_ALLOWED_USERS`	Yes	—	Comma-separated user ID allowlist
`HASHEE_SERVER_URL`	No	`https://api.hashee.ai`	API base URL
`HASHEE_CONNECTION_MODE`	No	`websocket`	`websocket` or `longpoll`

3. Start Claude Code

Claude Code will automatically detect the MCP server and make the Hashee tools available.

Channel Push

Incoming Hashee messages are pushed to Claude Code as channel notifications. Only messages from users in the HASHEE_ALLOWED_USERS list are forwarded.

The plugin sanitizes content (XML-escapes to prevent channel injection) and truncates at 10,000 characters.

Permission Relay

The plugin supports bidirectional permission relay between Claude Code and the Hashee app:

Claude Code requests a permission (e.g., to execute a command)
The plugin sends an Artifact to the Hashee app with approve/deny buttons
The user responds in the Hashee app
The decision is relayed back to Claude Code

Requests auto-deny after 60 seconds.

Security

Layer	Mechanism
Sender allowlist	Only `HASHEE_ALLOWED_USERS` can push to channel
XML sanitization	Prevents channel injection attacks
Content truncation	10,000 character limit
File path restriction	`send_file` only allows files within the working directory
Permission timeout	Auto-deny after 60 seconds
Agent auth	Bearer token on all REST calls

Next Steps

Codex Plugin — OpenAI Codex integration
OpenClaw Plugin — OpenClaw integration
SDK Getting Started — Build a standalone agent