The idea: keep the agents, throw away the orchestrator

Most setups do it the other way around. You keep one orchestrator running,
and it spawns temporary sub-agents per task — the agents are the disposable
part.

Hephaestus flips that. Specialist agents are packaged, versioned, and stored
in a hub, either local or shared. When a task comes in, a router picks the
right specialists out of that hub and assembles a temporary orchestrator
just for that task. It runs until the work is verified, and then the
orchestrator gets thrown away. The agents stick around for next time — and
across machines. I build an agent at home, upload it, and pull it down on my
work laptop instead of rebuilding it from scratch.

What's actually in it

A few pieces I added because I kept getting burned by the same problems:

Agents that lie about being done. Every execution goes through a gate
that blocks a "success" report until deterministic checks actually pass. It
reports back one of three states — verified, unverified, or blocked —
never just "done."

Agents built from a vague prompt that fall apart on edge cases. Before
compiling an agent, a short interview step scores how ambiguous the request
is across four axes — goal, constraints, scope, context — and won't build
until that's resolved. If your prompt is already clear, it skips the
interview entirely.

Routing that's a coin flip. Every agent ships a routing card with
explicit triggers and anti-triggers. Routing is a deterministic lookup
against those cards, not an LLM guessing which agent fits. Every routing
decision writes a plain-text receipt you can diff and commit.

Vendor lock-in. Each agent can run on a different model independently —
mix Claude, GPT, DeepSeek, GLM, or a local model through Ollama in the same
task force. If you're cost-sensitive, this matters a lot in practice: you
can run the whole thing on local models and pay nothing per token, or reserve
the expensive model calls for just the one specialist that needs them.

Try it

agentlas-ai / Hephaestus

Agent OS: keep specialist agents in a hub, spin up a temporary orchestrator per task. Local-first, works with any model.

Hephaestus — Model-Agnostic Agent OS

Stop building and configuring a new agent for every task. Hephaestus keeps specialist agents in a hub and spins up a temporary orchestrator per task.
Local-first, works with any model — Claude Code, Codex, Gemini, Cursor, and local models

English · Korean · 中文 · 日本語 · हिन्दी

_{Specialists pulled from the hub, assembled into a temporary task force, and routed live over MCP — no per-task agent setup.}

Quickstart

curl -fsSL https://raw.githubusercontent.com/agentlas-ai/Hephaestus/main/scripts/install-all-runtimes.sh | bash

This installs the neutral runner and registers command adapters for Claude Code, Codex, Gemini CLI, Antigravity, and Cursor. Prefer a plugin, a manual copy, or letting your AI install it for you? See All Install Methods.

The Agent OS Era · Quickstart · All Install Methods · Command Surface · New in v1.1.0 · Subsystems · Enterprise Operations · System Packaging · Docs Registry · Desktop Shell