Agent Hub

Agent-Hub

Your agents. Your providers. One local hub.

A local runtime for Claude Code and Codex.

Manage providers in one place, route Claude Code across Anthropic and
OpenAI-compatible APIs, and launch isolated Codex sessions without rewriting
your normal configuration.

English · 简体中文

Get started · Why Agent-Hub · Architecture · Changelog

Agent-Hub is a local runtime that owns the layer *around* your coding agent:
providers, sessions, routing, protocol translation and observability.

Providers live in Hub's own database. Nothing else needs to be running or
installed for normal use.

Provider management and session launching currently have separate terminal
entrypoints. See current boundaries before choosing an
interface.

Why Agent-Hub

Coding agents already own reasoning, tools, permissions and execution. Agent-Hub
focuses on the operational layer around them, so changing providers or routing
does not require replacing the agent loop or rewriting global configuration.

- Keep agents native. Claude Code and Codex remain their normal upstream
  runtimes; Agent-Hub sits beside them rather than forking their execution model.
- Manage providers once. Provider state lives in one local store and can be
  reused across isolated sessions.
- Translate only where needed. Claude Code can route through Anthropic-native
  or OpenAI-compatible upstreams; Codex keeps its native runtime and isolated
  CODEX_HOME path.
- Stay local-first. The control plane is local, credentials are redacted from
  normal responses, and complete message payloads are not persisted in gateway
  logs.

Where Agent-Hub fits

The runtime boundary is intentionally asymmetric: Claude Code can use the local
gateway and protocol bridge, while Codex stays close to its native configuration
model.

The two…