Local multi-agent manager

One desktop app for every coding agent

Agent Hub installs, connects, syncs Skills, and tracks usage for Claude Code, Codex, Kimi, Grok, and more — on your machine. Data stays local by default.

Windows / macOS / LinuxMIT licensedLocal-first
Agent Hub
Dashboard: status, token trend, and cost

More agents, more scattered work

Common problems when several coding agents live on the same machine

Installs stay scattered

Node, npm, Git, and each vendor installer live in different places. Upgrades and uninstalls are easy to lose track of.

Environment drift

Logins do not reuse

Subscriptions and API keys sit in each tool's config. Attaching the same login to another agent means setting it up again.

Repeated credentials

Skills copied by hand

The same skill has to be copied into every agent directory. One edit is easy to miss in the others.

Skills out of sync

Usage is hard to see

Tokens and cost sit in vendor logs. You cannot compare trend or spend in one view.

Opaque cost

One login, attached to more than one tool

What you store is an API key or a subscription login. Attaching it to another agent is only one of three methods. Prefer writing config. Local forwarding is the fallback.

Direct

Edit config only

The key already speaks the target tool's protocol. Agent Hub fills in the URL and model. No extra process.

Use this login

Write the login the other tool already understands

The target already knows this subscription. Agent Hub writes it into the place that tool expects. Still no extra process.

Local route

Local forwarding (fallback)

Only when the protocols do not match and this conversion is tested. The tool talks to 127.0.0.1. The real login stays in Agent Hub.

Daily work in one window

Filter by agent, navigate by job — not a separate island per tool

Dashboard

Agent status, token trends, cost estimates, and parse health. Installed agents can connect or switch from here.

Agents

Detect Node / npm / Git first, then install, upgrade, or remove each agent.

Connections

Official logins and API keys, then attach a login to another coding tool. Prefer config writes over local forwarding.

Skills

~/.agents/skills/ is the shared source of truth, synced out to each agent.

Chat

Call one or more local agents from the desktop app, with a streaming process view.

Projects

Browse and summarize local projects and sessions across agents.

MCP

Read-only summary of each agent's local MCP config. Injecting still happens in the tools themselves.

CLI

doctor, env, agent, skill, usage, backup, and more — same core as the GUI.

Agents we adapt today

Capabilities are uneven. Trust in-app detection and the capability matrix, not a brochure.

Claude CodeCodexKimiGrokPiWorkBuddyCursor AgentDeepSeek Harness

Local data by default

Credentials, sessions, and usage stay on your computer. They do not go through Chenshi's cloud.

On-disk paths

App data lives in ~/.agenthub/. Live config is backed up before a switch.

Usage is local-read

Usage parses local sessions or logs. No request interception, no upload.

Secrets masked

The UI, CLI, and logs do not echo full keys in the clear.

Product screens

Dashboard, Agents, Skills, and Chat are the four surfaces people use every day

Dashboard: status, token trend, and cost
Dashboard: status, token trend, and cost
Agents: fix the runtime, then install
Agents: fix the runtime, then install
Skills: one shared library, synced out
Skills: one shared library, synced out
Chat: talk to local agents on the desktop
Chat: talk to local agents on the desktop

FAQ

A desktop manager for coding agents — not custom Dify / N8N workflow delivery

Q: How is Agent Hub different from AI workflow?

A: Agent Hub is a desktop app that manages coding agents on an engineer's machine. AI workflow is our custom Dify / N8N delivery for business processes. One is local tooling; the other is enterprise automation.

Q: Do you upload our code or chats?

A: No. Default data stays in ~/.agenthub/ on the machine. Usage only reads local logs. There is no Chenshi cloud account sync, and we do not intercept or upload your requests.

Q: Which agents are supported?

A: Built-in adapters cover Claude Code, Codex, Kimi, Grok, Pi, WorkBuddy, Cursor Agent, and DeepSeek Harness. What each one can do differs — trust in-app detection. Cursor is the public Agent CLI only.

Q: How do we get it? Can we deploy inside a company network?

A: GitHub Releases ships Windows, macOS, and Linux builds under MIT. For internal distribution, private adapters, or training, use the consulting form.

Need a custom workflow instead?

If you want custom Dify / N8N workflows, document processing, or process automation, that is AI workflow — not this desktop app.

See AI workflow

Need internal distribution or extra adapters?

Individual developers can download from the button above. For internal distribution, private adapters, or wiring into an existing engineering process, use the consulting form.