Nexus MCP

An MCP server that enables AI models to invoke AI CLI agents (Codex, Claude Code, OpenCode) as tools. Provides parallel execution, automatic retries with exponential backoff, JSON-first response parsing, 10 discoverable prompt templates, model tier classification, and persistent preferences through seven MCP tools, four MCP resources, and ten MCP prompts.

Use Cases

Nexus MCP is useful whenever a task benefits from querying multiple AI agents in parallel rather than sequentially:

• Research & summarization — fan out a topic to multiple agents, then synthesize their responses into a single summary with diverse perspectives
• Code review — send different files or review angles (security, correctness, style) to separate agents simultaneously
• Multi-model comparison — prompt the same question to different models and compare outputs side-by-side for quality or consistency
• Bulk content generation — generate multiple test cases, translations, or documentation pages concurrently instead of one at a time
• Second-opinion workflows — get independent answers from separate agents before making a decision, reducing single-model bias

Features

• Parallel execution — batch_prompt fans out tasks with asyncio.gather and a configurable semaphore (default concurrency: 3)
• Automatic retries — exponential backoff with full jitter for transient errors (HTTP 429/503)
• Output handling — JSON-first parsing, brace-depth fallback for noisy stdout, temp-file spillover for outputs exceeding 50 KB
• Execution modes — default (safe, no auto-approve), yolo (full auto-approve)
• CLI detection — auto-detects binary path, version, and JSON output capability at startup
• Persistent preferences — set defaults for execution mode, model, retries, output limit, and timeout; preferences persist across MCP sessions via the backing store (MemoryStore default, FileTreeStore/RedisStore for restart persistence)
• Prompt templates — 10 discoverable workflow scaffolds (code review, debug, research, implement feature, etc.) via list_prompts/get_prompt; each returns structured messages with expert framing the client can use or ignore
• Model tier classification — heuristic-based model classification into quick/standard/thorough tiers; clients can override with sampling or live benchmarks. The nexus://runners resource includes tier data per model
• Tool timeouts — configurable safety timeout (default 15 min) cancels long-running tool calls to prevent the server from blocking indefinitely
• Client-visible logging — runner events (retries, output truncation, error recovery) are sent to MCP clients via protocol notifications, not just server stderr
• Elicitation — interactive parameter resolution via MCP elicitation; disambiguates missing CLI, offers model selection, confirms YOLO mode, and prompts for elaboration on vague prompts. Auto-detects client support and skips gracefully when unavailable. Suppression flags prevent repeat prompts within a session
• Benchmark data sources — server instructions include URLs for Artificial Analysis, OpenRouter, Chatbot Arena, and LLM Stats so clients can fetch live model benchmarks without API keys
• Extensible — implement build_command + parse_output, register in RunnerFactory

Installation

Run with uvx (recommended)

uvx nexus-mcp

uvx installs the package in an ephemeral virtual environment and runs it — no cloning required.

To check the installed version:

uvx nexus-mcp --version

To update to the latest version:

uvx --reinstall nexus-mcp

Claude Desktop (~/Library/Application Support/Claude/claude_desktop_config.json on macOS):

{
  "mcpServers": {
    "nexus-mcp": {
      "command": "uvx",
      "args": ["nexus-mcp"],
      "env": {
        "NEXUS_CODEX_MODEL": "gpt-5.2",
        "NEXUS_CODEX_MODELS": "gpt-5.4,gpt-5.4-mini,gpt-5.3-codex,gpt-5.2-codex,gpt-5.2,gpt-5.1-codex-max,gpt-5.1-codex-mini",
        "NEXUS_CLAUDE_MODEL": "claude-sonnet-4-6",
        "NEXUS_CLAUDE_MODELS": "claude-sonnet-4-6,claude-haiku-4-5-20251001",
        "NEXUS_OPENCODE_MODEL": "ollama-cloud/kimi-k2.5",
        "NEXUS_OPENCODE_MODELS": "ollama-cloud/glm-5,ollama-cloud/kimi-k2.5,ollama-cloud/qwen3-coder-next,ollama-cloud/minimax-m2.5,ollama/gemini-3-flash-preview"
      }
    }
  }
}

Cursor (.cursor/mcp.json in your project or ~/.cursor/mcp.json globally):

{
  "mcpServers": {
    "nexus-mcp": {
      "command": "uvx",
      "args": ["nexus-mcp"],
      "env": {
        "NEXUS_CODEX_MODEL": "gpt-5.2",
        "NEXUS_CODEX_MODELS": "gpt-5.4,gpt-5.4-mini,gpt-5.3-codex,gpt-5.2-codex,gpt-5.2,gpt-5.1-codex-max,gpt-5.1-codex-mini",
        "NEXUS_CLAUDE_MODEL": "claude-sonnet-4-6",
        "NEXUS_CLAUDE_MODELS": "claude-sonnet-4-6,claude-haiku-4-5-20251001",
        "NEXUS_OPENCODE_MODEL": "ollama-cloud/kimi-k2.5",
        "NEXUS_OPENCODE_MODELS": "ollama-cloud/glm-5,ollama-cloud/kimi-k2.5,ollama-cloud/qwen3-coder-next,ollama-cloud/minimax-m2.5,ollama/gemini-3-flash-preview"
      }
    }
  }
}

Claude Code (CLI):

claude mcp add nexus-mcp \
  -e NEXUS_CODEX_MODEL=gpt-5.2 \
  -e NEXUS_CODEX_MODELS=gpt-5.4,gpt-5.4-mini,gpt-5.3-codex,gpt-5.2-codex,gpt-5.2,gpt-5.1-codex-max,gpt-5.1-codex-mini \
  -e NEXUS_CLAUDE_MODEL=claude-sonnet-4-6 \
  -e NEXUS_CLAUDE_MODELS=claude-sonnet-4-6,claude-haiku-4-5-20251001 \
  -e NEXUS_OPENCODE_MODEL=ollama-cloud/kimi-k2.5 \
  -e NEXUS_OPENCODE_MODELS=ollama-cloud/glm-5,ollama-cloud/kimi-k2.5,ollama-cloud/qwen3-coder-next,ollama-cloud/minimax-m2.5,ollama/gemini-3-flash-preview \
  -- uvx nexus-mcp

Generic stdio config (any MCP-compatible client):

{
  "command": "uvx",
  "args": ["nexus-mcp"],
  "transport": "stdio",
  "env": {
    "NEXUS_CODEX_MODEL": "gpt-5.2",
    "NEXUS_CLAUDE_MODEL": "claude-sonnet-4-6",
    "NEXUS_OPENCODE_MODEL": "ollama-cloud/kimi-k2.5"
  }
}

All env keys are optional — see Configuration for the full list.

Prerequisites:

• Python 3.13+ (download)
• uv dependency manager (install guide)
  bash

  curl -LsSf https://astral.sh/uv/install.sh | sh
  

Optional (for integration tests):

• Codex — check with codex --version
• Claude Code — check with claude --version
• OpenCode — check with opencode --version

Claude Code note: Nexus invokes Claude Code non-interactively via claude -p. Anthropic says claude -p and Agent SDK usage draw from separate monthly Agent SDK credits starting 2026-06-15, while interactive Claude Code usage remains on plan usage limits: https://support.claude.com/en/articles/15036540-use-the-claude-agent-sdk-with-your-claude-plan

Note: Integration tests are optional. Unit tests run without CLI dependencies via subprocess mocking.

# 1. Clone the repository
git clone <repository-url>
cd nexus-mcp

# 2. Install dependencies
uv sync

# 3. Install pre-commit hooks (runs linting/formatting on commit)
uv run pre-commit install

# 4. Verify installation
uv run pytest                    # Run tests
uv run mypy src/nexus_mcp        # Type checking
uv run ruff check .              # Linting

# 5. Run the MCP server
uv run python -m nexus_mcp

⚠️ Experimental — This integration has not been validated end-to-end by the maintainer. Expect rough edges in setup, auth, and tool exposure. The MCP tools surfaced from upstream OpenCode track the upstream project and may change without notice. Feedback and bug reports are welcome.

Run an isolated OpenCode server for HTTP-based agent execution alongside the CLI runner. Provides session management, file search, permissions, and 38 additional MCP tools when the server is healthy.

Quick start:

1. Copy .env.example to .env and set PROJECT_DIR to your project path:
   bash

   cp .env.example .env
   # Edit .env: set PROJECT_DIR=/path/to/your/project
   

2. Start the server:
   bash

   docker compose up -d
   

3. Authenticate with your provider:
   bash

   docker exec -it opencode-server opencode auth login
   

4. Verify the server is healthy:
   bash

   curl -u opencode:nexus http://localhost:4096/global/health
   

The server binds to 127.0.0.1 (localhost only) by default for security. See docs/opencode-server-setup.md for the full guide including remote access, multi-project setup, and network security.

Usage

Once nexus-mcp is configured in your MCP client, your AI assistant automatically sees its tools. The reliable trigger is explicitly asking for output from an external AI agent (e.g. Codex, Claude Code, OpenCode). Generic "do this in parallel" prompts may be handled by the host AI's own capabilities instead. The cli parameter is optional — if omitted and the client supports MCP elicitation, the server will ask which runner to use. The server provides runner metadata (names, models, availability, execution modes) in its connection instructions — no discovery call needed. The cli parameter includes a JSON schema enum listing valid runner names.

Fan out a research question (batch_prompt)

You say: "Get perspectives from Codex, Claude Code, and OpenCode on transformer architectures."

{
  "tasks": [
    { "cli": "codex", "prompt": "Summarize the key findings of the Attention Is All You Need paper", "label": "codex-summary" },
    { "cli": "claude", "prompt": "What are the main limitations of transformer architectures?", "label": "claude-limitations" },
    { "cli": "opencode", "prompt": "List 3 real-world applications of transformers beyond NLP", "label": "opencode-applications" }
  ]
}

Code review from multiple angles (batch_prompt)

You say: "Have Codex, Claude Code, and OpenCode each review this diff in parallel."

{
  "tasks": [
    { "cli": "codex", "prompt": "Review this diff for security vulnerabilities:\n\n<paste diff>", "label": "codex-security-review" },
    { "cli": "claude", "prompt": "Review this diff for correctness and edge cases:\n\n<paste diff>", "label": "claude-correctness-review" },
    { "cli": "opencode", "prompt": "Review this diff for style and maintainability:\n\n<paste diff>", "label": "opencode-review" }
  ]
}

Single-agent prompt

You say: "Ask Codex to explain the difference between TCP and UDP."

{ "cli": "codex", "prompt": "Explain the difference between TCP and UDP in simple terms", "model": "gpt-5.2" }

Elicitation (server picks the runner)

You say: "Explain the CAP theorem using one of the available agents."

{ "prompt": "Explain the CAP theorem in simple terms" }

If the client supports MCP elicitation, the server asks which runner to use. Pass "elicit": false to skip.

Persistent preferences

You say: "Use YOLO mode with Codex from now on."

{ "execution_mode": "yolo", "model": "gpt-5.2", "max_retries": 5 }

Subsequent calls inherit these settings. Preferences persist across MCP sessions until explicitly cleared.

Fallback chain: explicit parameter → saved preference → per-runner env → global env → hardcoded default.

MCP Tools

All prompt tools run as background tasks — they return a task ID immediately so the client can poll for results, preventing MCP timeouts for long operations (e.g. YOLO mode: 2–5 minutes).

batch_prompt

Task object fields:

Note: elicit is a batch-level parameter. When enabled, the server runs a single upfront elicitation pass across all tasks rather than prompting per-task.

prompt

Same parameters as a single task object in batch_prompt, plus elicit (batch-level in batch_prompt, per-call here).

set_preferences

get_preferences / clear_preferences

get_preferences — no parameters, returns all fields (null when unset). clear_preferences — no parameters, resets all to null. Does not clear model tiers.

set_model_tiers

Persists tier classifications. Clients typically call once via sampling or benchmark fetch.

get_model_tiers

No parameters. Returns saved tiers as dict[str, str], or {} if none saved.

Managing Preferences

MCP Prompts

Nexus MCP provides 10 discoverable prompt templates that clients can browse via list_prompts() and render via get_prompt(name, args). Each prompt returns structured messages with expert framing — the client decides how (or whether) to use them.

Design principle: Server informs, client decides. Prompts provide the scaffold (role, structure, methodology); the client decides runner, model, depth, and orchestration. Prompts are completely optional — existing prompt/batch_prompt tools work exactly as before.

# 1. Client discovers available prompts
list_prompts() → sees "code_review", "debug", "compare_models", etc.

# 2. Client renders a prompt with arguments
get_prompt("code_review", {file: "src/auth.py", instructions: "security vulnerabilities"})

# 3. Server returns structured messages
→ PromptResult(
    messages=[
      Message("You are a senior code reviewer...", role="assistant"),
      Message("Review the file `src/auth.py`...\nFocus: security vulnerabilities\n...", role="user"),
    ],
    description="Code review of src/auth.py"
  )

# 4. Client feeds messages into prompt/batch_prompt with chosen runner+model
prompt(cli="claude", prompt=<rendered messages>)

MCP Resources

Read-only data endpoints that clients query for runner metadata, configuration, and preferences.

Models in nexus://runners include tier data: {"name": "gpt-5.4-mini", "tier": "quick"}. Tiers are quick (fast/cheap), standard (balanced), or thorough (max quality). Models with only heuristic tiers appear in unclassified_models — calling set_model_tiers moves them out.

Before set_model_tiers — all tiers are heuristic guesses, all models are unclassified:

{
  "models": [
    {"name": "gpt-5.1-codex-max", "tier": "thorough"},
    {"name": "gpt-5.4-mini", "tier": "quick"},
    {"name": "claude-sonnet-4-6", "tier": "standard"}
  ],
  "unclassified_models": ["gpt-5.1-codex-max", "gpt-5.4-mini", "claude-sonnet-4-6"]
}

After set_model_tiers — saved tiers replace heuristics, classified models leave the list:

{
  "models": [
    {"name": "gpt-5.1-codex-max", "tier": "thorough"},
    {"name": "gpt-5.4-mini", "tier": "quick"},
    {"name": "claude-sonnet-4-6", "tier": "standard"}
  ],
  "unclassified_models": []
}

Global Environment Variables

Per-Runner Environment Variables

Pattern: NEXUS_{AGENT}_{KEY} (agent name uppercased). Per-runner values override global values.

Valid {AGENT} values: CLAUDE, CODEX, OPENCODE, OPENCODE_SERVER

Invalid per-runner values are silently ignored (the global or hardcoded default is used instead).

Testing

This project follows Test-Driven Development (TDD) with strict Red→Green→Refactor cycles.

# Run all tests
uv run pytest

# Run with coverage report
uv run pytest --cov=nexus_mcp --cov-report=term-missing

# Run specific test types
uv run pytest -m integration           # Integration tests (requires CLIs)
uv run pytest -m "not integration"     # Unit tests only
uv run pytest -m "not slow"            # Skip slow tests

# Run specific test file
uv run pytest tests/unit/runners/test_codex.py

Test markers:

• @pytest.mark.integration — requires real CLI installations
• @pytest.mark.slow — tests taking >1 second

Code Quality

All quality checks run automatically via pre-commit hooks. Run manually:

# Lint and format
uv run ruff check .              # Check for issues
uv run ruff check --fix .        # Auto-fix issues
uv run ruff format .             # Format code

# Type checking (strict mode)
uv run mypy src/nexus_mcp

# Run all pre-commit hooks manually
uv run pre-commit run --all-files

Adding Dependencies

uv add <package>              # Production dependency
uv add --dev <package>        # Development dependency
uv sync                       # Sync environment after changes

Tool Configuration

• Ruff: line length 100, 17 rule sets (E/F/I/W + UP/FA/B/C4/SIM/RET/ICN/TID/TC/ISC/PTH/TD/NPY) — pyproject.toml → [tool.ruff]
• Mypy: strict mode, all type annotations required — pyproject.toml → [tool.mypy]
• Pytest: asyncio_mode = "auto", no @pytest.mark.asyncio needed — pyproject.toml → [tool.pytest.ini_options]
• Pre-commit: ruff-check, ruff-format, mypy, trailing-whitespace, end-of-file-fixer — .pre-commit-config.yaml

Python 3.13+ Syntax

• type keyword for type aliases: type AgentName = str
• Union syntax: str | None (not Optional[str])
• match statements for complex conditionals
• NO from __future__ import annotations

Project Structure

nexus-mcp/
├── src/nexus_mcp/
│   ├── __main__.py         # Entry point
│   ├── server.py           # FastMCP server + tools + prompt registration
│   ├── types.py            # Pydantic models
│   ├── exceptions.py       # Exception hierarchy
│   ├── config.py           # Environment variable config
│   ├── store.py            # Persistent backing store access (preferences + tiers)
│   ├── tiers.py            # Heuristic model tier classification
│   ├── elicitation.py      # ElicitationGuard — interactive parameter resolution
│   ├── resources.py        # MCP resources (runners, config, preferences)
│   ├── process.py          # Subprocess wrapper
│   ├── parser.py           # JSON→text fallback parsing
│   ├── cli_detector.py     # CLI binary detection + version checks
│   ├── prompts/
│   │   ├── __init__.py     # register_prompts(mcp) entry point
│   │   ├── analysis.py     # code_review, debug, quick_triage, research, second_opinion
│   │   ├── generation.py   # implement_feature, refactor, bulk_generate
│   │   ├── testing.py      # write_tests
│   │   └── comparison.py   # compare_models
│   └── runners/
│       ├── base.py         # Protocol + ABC
│       ├── factory.py      # RunnerFactory
│       ├── claude.py       # ClaudeRunner
│       ├── codex.py        # CodexRunner
│       ├── opencode.py     # OpenCodeRunner
│       └── opencode_server.py # OpenCode server runner
├── tests/
│   ├── unit/               # Fast, mocked tests
│   │   └── prompts/        # Prompt template tests
│   ├── e2e/                # End-to-end MCP protocol tests
│   ├── integration/        # Real CLI tests
│   └── fixtures.py         # Shared test utilities
├── .github/
│   └── workflows/          # CI, security, dependabot
├── pyproject.toml          # Dependencies + tool config
└── .pre-commit-config.yaml # Git hooks configuration

Releases

Stable releases are cut by running the Tag Release workflow from the Actions tab and choosing a bump (auto infers it from Conventional Commits since the last tag). Pre-releases are tagged manually. See RELEASE.md for the full maintainer workflow, recovery steps, and notes on server.json placeholder fields.

License

MIT