MCP Server details - Claude Octopus

The Claude Octopus MCP server exposes workflows as tools that any MCP-compatible client can consume. This enables integration with OpenClaw, Claude Desktop, Cursor, and other MCP clients.

Model Context Protocol overview

The Model Context Protocol (MCP) is an open protocol for connecting AI assistants to external tools and data sources. It provides:

Standardized tool interface — Uniform tool registration and invocation
Transport-agnostic — Works over stdio, HTTP, WebSocket
Security model — Sandboxed execution, permission controls
Type safety — JSON Schema validation for parameters

Claud Octopus implements an MCP server that delegates to orchestrate.sh, preserving exact behavioral parity with the Claude Code plugin.

Server architecture

Component diagram

┌─────────────────────────────────────────────────────────┐
│                     MCP Client                          │
│           (OpenClaw, Claude Desktop, etc.)              │
└────────────────────────┬────────────────────────────────┘
                         │ MCP Protocol (stdio/HTTP)
                         ▼
┌─────────────────────────────────────────────────────────┐
│                  MCP Server (Node.js)                   │
│  ┌──────────────────────────────────────────────────┐   │
│  │  Tool Registry (11 tools)                        │   │
│  │  - octopus_discover, octopus_define, etc.        │   │
│  └──────────────────────┬───────────────────────────┘   │
│                         │                               │
│  ┌──────────────────────▼───────────────────────────┐   │
│  │  Execution Engine                                 │   │
│  │  - Parameter validation (Zod)                    │   │
│  │  - Environment isolation                         │   │
│  │  - Security checks                               │   │
│  └──────────────────────┬───────────────────────────┘   │
└─────────────────────────┼───────────────────────────────┘
                          │ execFile()
                          ▼
┌─────────────────────────────────────────────────────────┐
│              orchestrate.sh (Bash)                      │
│  - Provider routing                                     │
│  - Quality gates                                        │
│  - Multi-agent orchestration                            │
└─────────────────────────────────────────────────────────┘

Startup sequence

Claude Code reads .mcp.json on plugin load
Spawns MCP server: node mcp-server/dist/index.js
Server initializes stdio transport
Registers 11 tools with MCP SDK
Loads skill metadata from .claude/skills/*.md
Sends server/initialized message to client
Client can now invoke tools

Configuration via .mcp.json

The .mcp.json file tells Claude Code how to launch the server:

{
  "mcpServers": {
    "octo-claw": {
      "command": "node",
      "args": ["./mcp-server/dist/index.js"],
      "env": {
        "CLAUDE_OCTOPUS_MCP_MODE": "true"
      }
    }
  }
}

Key points:

Server name: octo-claw (arbitrary identifier)
Command: node with path to compiled TypeScript
Environment: CLAUDE_OCTOPUS_MCP_MODE=true signals MCP execution context
Transport: stdio (stdin/stdout) — no network ports required

11 exposed tools

The MCP server registers these tools (see full API in OpenClaw integration):

Workflow tools (8)

Tool	Phase	Parameters
`octopus_discover`	Probe	`prompt`
`octopus_define`	Grasp	`prompt`
`octopus_develop`	Tangle	`prompt`, `quality_threshold?`
`octopus_deliver`	Ink	`prompt`
`octopus_embrace`	All 4 phases	`prompt`, `autonomy?`
`octopus_debate`	Debate	`question`, `rounds?`, `style?`, `mode?`
`octopus_review`	Review	`target`
`octopus_security`	Security	`target`

Introspection tools (2)

Tool	Purpose
`octopus_list_skills`	List all 44 available skills
`octopus_status`	Check provider availability

IDE integration tool (1)

Tool	Purpose
`octopus_set_editor_context`	Inject editor state (file, selection, cursor)

Connecting MCP clients

Claude Desktop

Add to ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or equivalent:

{
  "mcpServers": {
    "claude-octopus": {
      "command": "node",
      "args": ["/path/to/claude-octopus/mcp-server/dist/index.js"]
    }
  }
}

Restart Claude Desktop. The tools appear in the tool palette.

OpenClaw

Install the OpenClaw extension (see OpenClaw integration):

cd your-openclaw-installation
npm install github:nyldn/claude-octopus#main

The extension auto-connects to the MCP server.

Cursor

Cursor supports MCP via its settings:

{
  "mcp.servers": [
    {
      "name": "claude-octopus",
      "command": "node",
      "args": ["/path/to/claude-octopus/mcp-server/dist/index.js"]
    }
  ]
}

Custom MCP client

Use the MCP SDK:

import { Client } from "@modelcontextprotocol/sdk/client/index.js";
import { StdioClientTransport } from "@modelcontextprotocol/sdk/client/stdio.js";

const transport = new StdioClientTransport({
  command: "node",
  args: ["/path/to/claude-octopus/mcp-server/dist/index.js"],
});

const client = new Client({
  name: "my-client",
  version: "1.0.0",
}, {
  capabilities: {},
});

await client.connect(transport);

// List available tools
const tools = await client.listTools();
console.log(tools);

// Invoke a tool
const result = await client.callTool({
  name: "octopus_discover",
  arguments: {
    prompt: "Research OAuth 2.0 patterns"
  }
});

console.log(result);

Tool execution flow

Example: `octopus_develop`

Client invokes tool:

{
  "name": "octopus_develop",
  "arguments": {
    "prompt": "Build user authentication",
    "quality_threshold": 80
  }
}

Server validates parameters (Zod schema):

const schema = z.object({
  prompt: z.string().describe("What to implement"),
  quality_threshold: z.number().min(0).max(100).default(75)
});

Server builds environment:

const env = {
  PATH: process.env.PATH,
  HOME: process.env.HOME,
  CLAUDE_OCTOPUS_MCP_MODE: "true",
  // Forward only allowed env vars
  ...(process.env.OPENAI_API_KEY && { OPENAI_API_KEY: process.env.OPENAI_API_KEY }),
};

Server executes orchestrate.sh:

const { stdout, stderr } = await execFileAsync(
  "/path/to/orchestrate.sh",
  ["tangle", "Build user authentication"],
  { env, timeout: 300_000 }
);

orchestrate.sh runs workflow:
- Detects available providers (Codex, Gemini, Claude)
- Routes to appropriate agents
- Runs quality gates
- Returns synthesis file

Server returns result:

{
  "content": [
    {
      "type": "text",
      "text": "[Tangle synthesis output...]"
    }
  ],
  "isError": false
}

Security model

Environment variable isolation

Only allowed environment variables are forwarded to orchestrate.sh:

const BLOCKED_ENV_VARS = new Set([
  "OCTOPUS_SECURITY_V870",      // Security hardening toggle
  "OCTOPUS_GEMINI_SANDBOX",     // Sandbox mode
  "OCTOPUS_CODEX_SANDBOX",      // Sandbox mode
  "CLAUDE_OCTOPUS_AUTONOMY",    // Autonomy level
]);

const env = {
  // Only essential system vars
  PATH: process.env.PATH,
  HOME: process.env.HOME,
  TMPDIR: process.env.TMPDIR,
  
  // AI provider keys (if set)
  ...(process.env.OPENAI_API_KEY && { OPENAI_API_KEY: process.env.OPENAI_API_KEY }),
  
  // Octopus config (filtered against blocklist)
  ...Object.fromEntries(
    Object.entries(process.env).filter(([k]) =>
      (k.startsWith("OCTOPUS_") || k.startsWith("CLAUDE_OCTOPUS_")) &&
      !BLOCKED_ENV_VARS.has(k)
    )
  ),
};

See source code:~/workspace/source/mcp-server/src/index.ts:53-96

API key sanitization

Error messages sanitize leaked API keys:

const sanitized = msg.replace(/[A-Za-z_]+KEY=[^\s]+/g, "[REDACTED]");

Path validation

IDE context paths reject traversal attempts:

for (const [label, value] of [["filename", filename], ["workspace_root", workspace_root]]) {
  if (value && /\.\.[\\/ ]/.test(value)) {
    return {
      content: [{ type: "text", text: `Error: ${label} cannot contain '..'` }],
      isError: true,
    };
  }
}

Selection size limit

Editor selections are truncated to 50KB:

const MAX_SELECTION_LENGTH = 50_000; // 50KB
const safeSel = selection && selection.length > MAX_SELECTION_LENGTH
  ? selection.slice(0, MAX_SELECTION_LENGTH)
  : selection;

Transport security

stdio transport (default) is scoped to the spawning process — no network exposure.

If switching to HTTP/SSE/WebSocket transport, add bearer token authentication.

Skill metadata

The server loads skill metadata from frontmatter:

async function loadSkillMetadata(): Promise<SkillMeta[]> {
  const skillsDir = resolve(PLUGIN_ROOT, ".claude/skills");
  const files = await readdir(skillsDir);
  
  for (const file of files) {
    if (!file.endsWith(".md")) continue;
    const content = await readFile(resolve(skillsDir, file), "utf-8");
    const frontmatterMatch = content.match(/^---\n([\s\S]*?)\n---/);
    if (!frontmatterMatch) continue;
    
    const fm = frontmatterMatch[1];
    const name = fm.match(/^name:\s*(.+)$/m)?.[1]?.trim();
    const description = fm.match(/^description:\s*["']?(.+?)["']?\s*$/m)?.[1]?.trim();
    
    skills.push({ name, description, file });
  }
  
  return skills;
}

See source code:~/workspace/source/mcp-server/src/index.ts:121-152

Debugging the MCP server

Enable debug logs

export OCTOPUS_DEBUG=true
node mcp-server/dist/index.js

Test with MCP Inspector

Use the MCP Inspector to test the server:

npx @modelcontextprotocol/inspector node mcp-server/dist/index.js

This opens a web UI for invoking tools and inspecting responses.

Check server logs

Server logs are written to:

Claude Code: ~/.claude/logs/mcp-server.log
OpenClaw: ~/.openclaw/logs/mcp-server.log

Common issues

Server fails to start

Symptoms: Client reports “MCP server not available”Causes:

Missing dependencies: Run npm install in mcp-server/
TypeScript not compiled: Run npm run build in mcp-server/
Wrong Node.js version: Requires Node.js ≥18

Fix:

cd mcp-server
npm install
npm run build
node dist/index.js  # Should not error

Tool invocation times out

Symptoms: Client times out waiting for responseCauses:

orchestrate.sh hung on provider call
Quality gate blocked execution
Large synthesis file being generated

Fix:

Check ~/.claude-octopus/logs/orchestrate.log
Increase timeout in client config
Enable debug mode: OCTOPUS_DEBUG=true

Environment variables not passing through

Symptoms: orchestrate.sh can’t find API keysCauses:

Env vars not set in MCP server environment
Blocked by security filter

Fix:

Set env vars in .mcp.json:

{
  "env": {
    "OPENAI_API_KEY": "sk-...",
    "GEMINI_API_KEY": "..."
  }
}

Check blocked vars list in mcp-server/src/index.ts:53

Source code reference

MCP server: mcp-server/src/index.ts in source code:~/workspace/source/mcp-server/src/index.ts
MCP config: .mcp.json in source code:~/workspace/source/.mcp.json
Skill schema: mcp-server/src/schema/skill-schema.json
Package config: mcp-server/package.json

OpenClaw integration — Installing and configuring OpenClaw extension
Workflows — Understanding workflow phases
Security — Security model and threat mitigation
Debugging guide — Troubleshooting workflows

Extension & Integration

Troubleshooting

Documentation Index

​Model Context Protocol overview

​Server architecture

​Component diagram

​Startup sequence

​Configuration via .mcp.json

​11 exposed tools

​Workflow tools (8)

​Introspection tools (2)

​IDE integration tool (1)

​Connecting MCP clients

​Claude Desktop

​OpenClaw

​Cursor

​Custom MCP client

​Tool execution flow

​Example: octopus_develop

​Security model

​Environment variable isolation

​API key sanitization

​Path validation

​Selection size limit

​Transport security

​Skill metadata

​Debugging the MCP server

​Enable debug logs

​Test with MCP Inspector

​Check server logs

​Common issues

​Source code reference

​Related documentation

Model Context Protocol overview

Server architecture

Component diagram

Startup sequence

Configuration via .mcp.json

11 exposed tools

Workflow tools (8)

Introspection tools (2)

IDE integration tool (1)

Connecting MCP clients

Claude Desktop

OpenClaw

Cursor

Custom MCP client

Tool execution flow

Example: `octopus_develop`

Security model

Environment variable isolation

API key sanitization

Path validation

Selection size limit

Transport security

Skill metadata

Debugging the MCP server

Enable debug logs

Test with MCP Inspector

Check server logs

Common issues

Source code reference

Related documentation