Hermes Global — Open-Source Multi-Agent Orchestration CLI

The white-box alternative to Sakana Fugu. Not a black-box router — a transparent AI team you can see, control, and customize.

---

What It Is

A CLI tool that spins up a team of 5 specialized AI agents — CEO, PM, Dev, CSO, and Ops — to tackle coding tasks collaboratively. You bring your own API keys. You see every decision. You control every model.

Why It Exists

Fugu proves multi-agent > single-agent (93.2% LiveCodeBench). But Fugu is:

• ❌ Closed-source — you can't see how routing decisions are made
• ❌ Expensive — $30/1M output tokens for Ultra tier
• ❌ Model-locked — you can't choose which models are used

Google Antigravity is promising but:

• ❌ Google-controlled — single vendor lock-in
• ❌ New (May 2026) — unproven at scale

Hermes Global is:

• ✅ Open source (MIT) — audit every line
• ✅ Model-free — use DeepSeek, GLM, GPT, Claude, or any OpenAI-compatible API
• ✅ Observable — every agent decision is logged and replayable
• ✅ Local-first — runs on your machine, not our servers
• ✅ Extensible — add your own agents, tools, and skills

---

Quick Start

Prerequisites

• Python 3.11+ (check with python --version)
• Hermes Agent installed (pip install hermes-agent or see hermes-agent.nousresearch.com)
• An API key from your preferred LLM provider

Installation

# Install from PyPI
pip install hermes-global

# Or with uv
uv tool install hermes-global

Initial Setup

# Interactive setup wizard — picks provider, creates 5-agent team
hermes-global init

The wizard will guide you through:

1. Choosing your model provider (OpenAI / Anthropic / DeepSeek / Zhipu / Custom)
2. Entering your API key
3. Configuring proxy settings (important for users in China)
4. Creating the 5 agent profiles
5. Initializing the kanban board

Run Your First Task

# Delegate a task to the AI team
hermes-global run "Build a REST API for user authentication"

# Simple tasks get fast, direct execution
hermes-global run "Fix typo in utils.py line 42"

# See what's happening with your team
hermes-global status

---

The 5-Role AI Team

Role	Icon	Responsibility
CEO	🧠	Strategic direction, GO/NO-GO decisions, scope management
PM	📋	Requirements clarification, spec writing, priority ordering
Dev	💻	Code implementation, TDD, bug fixing
CSO	🔒	Security audit, CVE scanning, secret detection
Ops	🚀	CI/CD, testing, release management, monitoring

Every agent has its own profile, model configuration, and skill set. The orchestrator automatically routes tasks through the appropriate agents based on task complexity and security requirements.

---

5 Dynamic Topologies

The system automatically selects the best execution pattern for your task:

Topology	Use Case	Pipeline
Direct ⚡	Simple fixes, typos	Dev only
Serial ➡️	Complex, security-critical	CEO→PM→Dev→CSO→Ops
Parallel ∥	Medium complexity	PM→(Dev∥CSO)→Ops
Tree 🌳	Multi-faceted tasks	CEO→parallel sub-pipelines
Decompose 🧩	Large, multi-module tasks	CEO→independent workstreams

You can also force a specific topology:

hermes-global run --topology direct "Rename variable x to count"
hermes-global run --topology serial "Refactor auth system to OAuth 2.0"

---

Full Observability

Every orchestration step is visible in real time:

📋 Task Analysis
  Task:      Add JWT login to the API
  Topology:  PARALLEL
  Reasoning: Medium-complexity task. Parallel Dev+CSO review for efficiency.
  Pipeline:  PM → DEV ∥ CSO → OPS

  ✅ [🧠 CEO] GO — approved (0.3s)
  ✅ [📋 PM]  Spec written (2.1s)
  ✅ [💻 DEV] Code implemented (8.4s)
  ✅ [🔒 CSO] Audit PASS — 0 CRITICAL (1.5s)
  ✅ [🚀 OPS] Tests passed, packaged (3.2s)

═══ Orchestration Complete ═══
  Topology: parallel
  Steps: 5 (5 ✅, 0 ❌)
  Total time: 15.5s
  Est. tokens: ~4,200

---

Model Freedom

Configure any provider per agent. Mix and match to optimize cost and quality:

# In ~/.hermes-global/config.yaml
agents:
  ceo:
    provider: anthropic
    model: claude-sonnet-4-20250514    # best reasoning
  dev:
    provider: deepseek
    model: deepseek-v4-pro             # best coding
  ops:
    provider: openai
    model: gpt-4o-mini                 # cheap for CI/CD

Supported providers: OpenAI, Anthropic, DeepSeek, Zhipu AI (GLM), and any OpenAI-compatible API (Ollama, vLLM, Groq, OpenRouter, etc.)

---

Commands

hermes-global init              # Interactive setup wizard
hermes-global run <task>        # Execute a task through the AI team
hermes-global status            # Show team status, kanban, capabilities
hermes-global --help            # Show all commands and options
hermes-global run --help        # Show run command options

---

Configuration

Configuration lives at ~/.hermes-global/config.yaml. See config_template.yaml for a fully documented example.

API keys are stored in ~/.hermes-global/.env (never committed to git):

DEEPSEEK_API_KEY=sk-...
OPENAI_API_KEY=sk-...
ANTHROPIC_API_KEY=sk-ant-...

---

Tech Stack

• Python 3.11+ / uv for package management
• Click for CLI
• Rich for terminal output
• 5 specialized agents (CEO/PM/Dev/CSO/Ops)
• 5 dynamic topologies (direct/serial/parallel/tree/decompose)
• Auto prompt engineering (4-layer: role/context/format/pitfalls)
• 3-level recursive self-correction (Dev↔CSO loop)
• Capability matrix-driven routing (learns who's good at what)

---

Comparison

	Hermes Global	Fugu	Antigravity
Open source	✅ MIT	❌	❌
Model freedom	✅ Any API	❌ Sakana only	❌ Google only
Observable	✅ Full tracing	❌ Black box	🟡 Partial
Pricing	Your API keys	$5-30/1M tokens	Free (for now)
Topologies	5 (direct/serial/parallel/tree/decompose)	Unknown	Parallel agents
Custom agents	✅ Yes	❌	❌
Local-first	✅	❌	❌

---

Roadmap

• P0: CLI (init/run/status) + 5 topologies + 5 roles + observability
• P1: Recursive self-correction + capability-matrix routing + auto prompt engineering
• P2: Kanban management CLI + completion scripts + JSON/Markdown export
• P3: Docker deployment + community skill marketplace + multi-language UI

---

Security

API keys are stored in ~/.hermes-global/.env. This file is plaintext — treat it like a password manager master key:

• Unix: Run chmod 600 ~/.hermes-global/.env after init
• Windows: Run icacls "%USERPROFILE%\.hermes-global\.env" /inheritance:r /grant:r "%USERNAME%:F"
• Never commit .env to git (it's in .gitignore by default)
• For production use, consider a secrets manager (Bitwarden, HashiCorp Vault)

Prompt injection warning: Hermes Global accepts natural-language task descriptions and delegates them to AI agents with code execution capabilities. For untrusted input, use the --sandbox flag (coming in P2) or run with a containerized backend.

Known framework dependencies: Hermes Global builds on Hermes Agent, which carries its own security posture. See Hermes Agent Security for the full threat model.

---

Contributing

Hermes Global is part of the Hermes Agent ecosystem by Nous Research.

• Issues: GitHub Issues
• Docs: hermes-agent.nousresearch.com/docs
• License: MIT

---

_{Built with ❤️ by the Hermes Agent community}