openclaw-cursor-brain is an OpenClaw plugin that turns Cursor Agent CLI into a fully-integrated, streaming AI backend via MCP. All OpenClaw plugin tools become natively accessible to Cursor — and vice versa.

Prerequisites

• Cursor IDE installed and launched at least once

• Cursor Agent CLI enabled: open Cursor → Cmd+Shift+P → type Install 'cursor' command → Enter. Or execute shell curl https://cursor.com/install -fsS | bash. (after this, cursor-agent or agent is available in the terminal)

• OpenClaw installed globally (npm i -g openclaw)

• Node.js ≥ 18

Quick Start

openclaw plugins install openclaw-cursor-brain # install & default config
openclaw gateway restart                       # restart gateway
openclaw cursor-brain doctor                   # verify

Model selection (optional): In a TTY, interactive model selection may run automatically after install. Otherwise run:

openclaw cursor-brain setup   # optional: choose primary & fallback models
openclaw gateway restart      # if you changed config

During setup (and upgrade), models are dynamically discovered from cursor-agent --list-models. Primary is single-select; fallbacks are multi-select (space to toggle), defaulting to all models in cursor's original order:

◆  Select primary model (↑↓ navigate, enter confirm)
│  ● auto              Auto
│  ○ opus-4.6-thinking Claude 4.6 Opus (thinking, cursor default)
│  ○ opus-4.6          Claude 4.6 Opus
│  ○ sonnet-4.6        Claude 4.6 Sonnet
│  ○ gpt-5.3-codex     GPT-5.3 Codex
│  …
└

◆  Select fallback models (space toggle, enter confirm, order follows list)
│  ◼ opus-4.6-thinking Claude 4.6 Opus (thinking, cursor default)
│  ◼ opus-4.6          Claude 4.6 Opus
│  ◼ sonnet-4.6        Claude 4.6 Sonnet
│  …
└

How It Works

flowchart TD
    subgraph Channels ["📱 Messaging Channels"]
        direction LR
        Feishu["Feishu"]
        Slack["Slack"]
        Web["Web / API"]
    end

    GW["OpenClaw Gateway"]

    subgraph Bridge ["🔄 Bidirectional Bridge"]
        direction LR
        subgraph ProxyBox ["⚡ Streaming Proxy :18790"]
            Proxy["OpenAI-compatible API<br/>Session auto-derive<br/>InstantResult · scriptHash auto-restart"]
        end
        subgraph MCPBox ["🔌 MCP Server (stdio)"]
            MCP["Tool Gateway<br/>Rich descriptions · Timeout + retry<br/>Auto-discovery"]
        end
    end

    Agent["🧠 cursor-agent CLI<br/>--stream-partial-output --resume"]

    subgraph Tools ["🛠️ Plugin Tools"]
        direction LR
        T1["feishu_doc"]
        T2["feishu_wiki"]
        T3["GitHub · DB · …"]
    end

    Channels -->|"User messages"| GW
    GW -->|"POST /v1/chat/completions"| Proxy
    Proxy -->|"spawn + stdin"| Agent
    Agent -->|"SSE stream"| Proxy
    Proxy -->|"Response"| GW
    GW -->|"Reply"| Channels

    Agent <-->|"MCP stdio"| MCP
    MCP -->|"POST /tools/invoke"| GW
    GW --> Tools

Two auto-configured paths:

1. Streaming Proxy — local OpenAI-compatible server (/v1/chat/completions) spawns cursor-agent with --stream-partial-output, streams text deltas as SSE in real-time. Includes tool call logging, instant result delivery, optional thinking forwarding, and script hash-based auto-restart on upgrade.
2. MCP Server — Cursor IDE calls OpenClaw tools via stdio, proxied to Gateway REST API with timeout/retry. Server instructions include rich tool descriptions (token extraction rules, action keys, parameter examples) to minimize unnecessary openclaw_skill calls.

Sessions are persisted to disk and reused via --resume for faster subsequent responses (context survives restarts). Session IDs are automatically derived from conversation metadata embedded in user messages (e.g., sender_id for DM, group_channel + topic_id for group chats), or can be passed explicitly via body fields (_openclaw_session_id, session_id) or HTTP headers (X-OpenClaw-Session-Id, X-Session-Id). New plugins are auto-discovered on gateway restart.

Bidirectional Enhancement

• OpenClaw gains Cursor AI — all channels (Slack, Feishu, Web, etc.) get Cursor's frontier models as the AI backend
• Cursor IDE gains OpenClaw tools — all plugin tools auto-registered as MCP tools, letting Cursor natively call Feishu, Slack, GitHub, databases, etc.
• Compound effect — a single agent session can read Slack, write code, push to GitHub, and notify on Feishu — no context switching

Features

• Zero config — install and restart; everything auto-configures
• Interactive model selection — setup/upgrade present all discovered models via @clack/prompts (single-select primary, multi-select ordered fallbacks)
• Dynamic model discovery — models auto-detected from cursor-agent --list-models, synced to OpenClaw on every gateway start
• Real-time streaming — --stream-partial-output for character-level text deltas; instant result by default (plugin config instantResult), with optional smart-chunked fallback
• Thinking forwarding — optionally stream LLM reasoning (plugin config forwardThinking): "off" drop; "content" (default) as markdown blockquote in message body (compatible with OpenClaw Feishu/Slack streaming cards); "reasoning_content" via standard field
• Rich tool descriptions — MCP server instructions include token extraction rules, exact action keys, and parameter examples from SKILL.md — reducing unnecessary openclaw_skill calls
• Tool call logging — proxy logs every tool invocation with name, arguments summary, duration, and call ID for diagnostics
• Tool auto-discovery — disk-based registration from SKILL.md at startup (no Gateway dependency); background verification for diagnostics; cached with 60s TTL
• Session auto-derive — session keys automatically derived from conversation metadata (sender/group/topic) in user messages; no explicit session ID required from Gateway
• Session persistence — cursor-agent sessions persisted to disk (~/.openclaw/cursor-sessions.json); explicit session IDs also accepted via body fields or HTTP headers (X-OpenClaw-Session-Id, X-Session-Id)
• Auto-restart on upgrade — proxy exposes a scriptHash in /v1/health; gateway compares it against the installed script hash and auto-restarts when code changes
• Three-layer fault tolerance — request-level: auto-clear stale session and retry once; process-level: self-exit after consecutive failures to trigger restart; gateway-level: exponential backoff auto-restart on crash (2s → 10s → 60s, resets after 5min stable)
• Version guard — upgrade command detects version comparison; downgrade or same-version requires confirmation
• Standalone proxy — streaming-proxy.mjs runs independently as OpenAI-compatible API (auto-detection, API key auth, CORS)
• Reliability — tool call timeout (60s), retry (2x), structured MCP errors (isError: true)
• Proxy management — proxy status/stop/restart/log commands for lifecycle control without restarting gateway
• Diagnostics — doctor (10+ checks), status, and proxy log commands
• Request safety — request body size limit (10 MB), per-request timeout, graceful client disconnect handling
• Cross-platform — macOS, Linux, Windows (port management, health checks)

Configuration

In openclaw.json under plugins.entries.openclaw-cursor-brain.config:

| Option | Type | Default | Description | | -------------- | ------ | ----------- | --------------------------- | | cursorPath | string | auto-detect | Path to cursor-agent binary | | outputFormat | string | auto-detect | "stream-json" or "json" | | proxyPort | number | 18790 | Streaming proxy port |

Primary and fallback models are not in plugin config; they are stored in agents.defaults.model (primary + fallbacks array) and in models.providers.cursor-local. Use openclaw cursor-brain setup or the upgrade flow to set them interactively.

Single source of truth: Proxy options are configured only in openclaw.json under plugins.entries.openclaw-cursor-brain.config. The proxy always reads from openclaw.json: when started by the gateway it uses OPENCLAW_CONFIG_PATH; when run standalone (node streaming-proxy.mjs) it defaults to ~/.openclaw/openclaw.json. No separate config file. Uninstall stops the proxy and removes any legacy cursor-proxy.json; upgrade preserves plugin config in openclaw.json.

| Variable | Default | Description | | --------------------------- | -------- | ------------------------------------ | | OPENCLAW_TOOL_TIMEOUT_MS | 60000 | MCP tool call timeout (ms) | | OPENCLAW_TOOL_RETRY_COUNT | 2 | MCP retries on transient errors | | CURSOR_PROXY_PORT | 18790 | Port (when running proxy standalone) | | CURSOR_PROXY_API_KEY | (none) | API key for standalone proxy auth |

CLI Commands

| Command | Description | | ---------------------------------------- | ------------------------------------------- | | openclaw cursor-brain setup | Configure MCP + interactive model selection | | openclaw cursor-brain doctor | Health check (10+ items) | | openclaw cursor-brain status | Show versions, config, models & tool count | | openclaw cursor-brain upgrade <source> | One-command upgrade + model selection | | openclaw cursor-brain uninstall | Full uninstall (configs + files) | | openclaw cursor-brain proxy | Show proxy status (PID, port, sessions) | | openclaw cursor-brain proxy stop | Stop the streaming proxy | | openclaw cursor-brain proxy restart | Restart proxy (detached) | | openclaw cursor-brain proxy log [-n N] | Show last N lines of proxy log (default 30) |

Standalone Streaming Proxy

The proxy works without OpenClaw, turning any Cursor into an OpenAI-compatible API:

node mcp-server/streaming-proxy.mjs
# with options:
CURSOR_PROXY_PORT=8080 CURSOR_PROXY_API_KEY=secret node mcp-server/streaming-proxy.mjs

curl http://127.0.0.1:18790/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"model":"auto","stream":true,"messages":[{"role":"user","content":"Hello!"}]}'

Endpoints: POST /v1/chat/completions, GET /v1/models, GET /v1/health (returns scriptHash for version detection). Supports API key auth, CORS, session reuse via body or headers.

~/.cursor/mcp.json

{
  "mcpServers": {
    "openclaw-gateway": {
      "command": "node",
      "args": ["<plugin-path>/mcp-server/server.mjs"],
      "env": {
        "OPENCLAW_GATEWAY_URL": "http://127.0.0.1:<port>",
        "OPENCLAW_GATEWAY_TOKEN": "<token>",
        "OPENCLAW_CONFIG_PATH": "~/.openclaw/openclaw.json"
      }
    }
  }
}

openclaw.json (after interactive selection)

{
  "agents": {
    "defaults": {
      "model": {
        "primary": "cursor-local/auto",
        "fallbacks": ["cursor-local/opus-4.6", "cursor-local/sonnet-4.6", "..."]
      }
    }
  },
  "models": {
    "providers": {
      "cursor-local": {
        "api": "openai-completions",
        "baseUrl": "http://127.0.0.1:18790/v1",
        "apiKey": "local",
        "models": [
          { "id": "auto", "name": "Auto" },
          { "id": "opus-4.6", "name": "Claude 4.6 Opus" },
          "..."
        ]
      }
    }
  }
}

| Problem | Fix | | ------------------------------------------------------------------ | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | | WARNING: dangerous code patterns (child_process / env+network) | OpenClaw scans plugins and may show this during install. This plugin legitimately uses child_process to run the Cursor agent and proxy, and reads env (e.g. CURSOR_PATH) to configure them. The plugin only passes a minimal env whitelist to the proxy child, not full process.env. You can ignore this warning for this plugin. | | Stuck after "Provider synced" | Old versions started proxy/timers during install; fixed in current release. Upgrade to latest. | | No model selection during install | Runs only in a TTY; otherwise run openclaw cursor-brain setup after install. | | Invalid config … source / unknown command 'cursor-brain' | Old versions wrote invalid source: "tarball". New version auto-fixes on register. If it still fails, edit ~/.openclaw/openclaw.json, set plugins.installs.openclaw-cursor-brain.source to "archive" (or "path" for local install), save, then run openclaw plugins install ./ again. | | plugins.allow / plugins.entries: plugin not found | Stale config references the uninstalled plugin; openclaw doctor --fix may not remove it. From the project dir run npm run uninstall, then openclaw plugins install ./. Or edit ~/.openclaw/openclaw.json: remove openclaw-cursor-brain from plugins.allow and delete plugins.entries.openclaw-cursor-brain. | | No API key found for provider "anthropic" | The default model is anthropic (needs API key), not cursor-local. Use local Cursor: run openclaw config set agents.defaults.model.primary "cursor-local/auto", or openclaw cursor-brain setup to pick cursor-local/auto, then openclaw gateway restart. If an agent overrides to anthropic, check openclaw config get agents.list and fix that agent’s model.primary. | | "Cursor Agent CLI not found" | Install Cursor and launch once, or set config.cursorPath | | Gateway error | Confirm gateway running (openclaw gateway status), check token | | Tools not appearing | Restart gateway, call openclaw_discover in Cursor | | Tool timeout | Set OPENCLAW_TOOL_TIMEOUT_MS=120000 | | Proxy not starting | openclaw cursor-brain proxy log to check; proxy restart to force start | | Cursor rate limit / 504 timeout / proxy keeps exiting | When Cursor is rate-limited, requests slow down. Defaults: 5 min request timeout, 5 min when degraded; health reports "degraded" only when consecutive failures ≥4 or consecutive timeouts ≥2, so Gateway does not restart proxy on a single timeout. Non-stream requests accumulate agent text output so partial content is returned if any was received before timeout. Channels like Feishu often have ~50–60s reply timeout; if Cursor is often slower, reduce concurrency or wait for rate limit to ease. | | Proxy not updating after upgrade | Gateway auto-restarts proxy when scriptHash differs; verify with curl http://127.0.0.1:18790/v1/health | | Slow batch responses | In plugin config set instantResult: false; then batch results are chunked at ~200 chars/s | | Context lost between messages | Check ~/.openclaw/logs/cursor-proxy.log for session=auto:dm:…(meta.auto) — if showing session=none(none), ensure Gateway embeds "Conversation info" metadata in user messages | | Context lost after restart | Sessions auto-persist to disk; use proxy restart (not gateway restart) to keep sessions | | Debug MCP server | OPENCLAW_GATEWAY_URL=... node mcp-server/server.mjs | | Debug tool calls | Check ~/.openclaw/logs/cursor-proxy.log for tool:start / tool:done entries |

openclaw-cursor-brain/
  index.ts                  # Plugin entry (register + CLI commands)
  openclaw.plugin.json      # Plugin metadata + config schema
  src/
    constants.ts            # Paths, IDs, output format types
    setup.ts                # Idempotent setup, model discovery, format detection
    doctor.ts               # Health checks (11 items)
  scripts/
    uninstall.mjs           # Uninstall: config + MCP + extension dir
  mcp-server/
    server.mjs              # MCP server (tool discovery, timeout/retry)
    streaming-proxy.mjs     # OpenAI-compatible streaming proxy

License

MIT