Agent Workbench

Five coding agents.
One window.

A local browser interface around the CLIs you already use.

Install with

npx agent-workbench

# try it once; about 60 MB the first time

npm install -g agent-workbench

agent-workbench# in your project’s folder

git clone https://github.com/cvelasquez/agent-workbench.git

cd agent-workbench && corepack enable pnpm && pnpm install && pnpm dev

  • MIT licence
  • 127.0.0.1 only
  • no network calls of its own

What does it add to my terminal?

Your CLI stays exactly as it is. Agent Workbench adds what a terminal alone doesn’t give you.

  1. claude codex opencode agy pi# unchanged
  2. tabs, one per agent, each in its own folder
  3. every past conversation, from all five CLIs
  4. the conversation as easy-to-read cards
  5. a context meter, in tokens, never money
  6. the files that changed, and their diff
  • 5# CLIs it works with
  • 9# interface languages
  • 0# network calls of its own
  • 1# command to try it

What sits around my CLI?

The terminal is still the terminal: everything you type goes straight to the CLI. The rest sits around it.

Show

Cards, changes and a diff, one tab per CLI, and a conversation continued with another CLI.

The main view: a conversation as cards, with a tab per agent and the project’s files beside it

The messages as cards, live, with tool calls folded. Answer the agent’s questions right there, depending on the CLI.

The Changes panel: branch and changed files, by group The diff of one of those files, in the same panel

The branch, the changed files and their diff. Open a new branch in its own folder so several agents work in parallel.

A file from the tree, previewed with syntax highlighting

The file tree with search and a highlighted preview. Edit a text file without overwriting the agent’s changes.

The Memory tab with the bridge installed for the five CLIs and the imported notes

Shared memory between the five CLIs, with the notes it imported. Optional, and off by default.

Can I see it working first?

The whole app in a six-minute tutorial, and one short per feature, none longer than 40 seconds.

A frame from the video tutorial: three branches, each in its own folder, with Open, See changes, Ask for a review and Merge into main Play the tutorial

The whole app in 6 minutes

With chapters, from the first tab to merging a branch. YouTube loads only when you press play.

More on the YouTube channel

One feature per short, with sound

  1. 5 agents. 1 window.

    Five agents, one window

    0:32

  2. 3 agents. Zero collisions.

    Three agents on three branches

    0:37

  3. Out of quota? Keep going.

    Out of quota? Continue with another CLI

    0:18

  4. Plan usage. At a glance.

    Plan usage at a glance

    0:25

  5. 3 AM reset? Prompt ready.

    Schedule the next prompt

    0:22

  6. Find any conversation.

    Search every conversation

    0:22

  7. Edit .env. Keep both changes.

    Edit a .env without overwriting the agent

    0:30

  8. Agent done. Phone knows.

    Your phone tells you it’s done

    0:28

Every short on YouTube

Does it touch my credentials?

No. It reads each CLI’s history and writes nothing in their folders. There’s no login screen.

What Agent Workbench reads from each CLI, and what it never opens
CLIReadsNever opens
Claude Code~/.claude/projects/, sessions/, settings.json.credentials.json or any token
Codex~/.codex/sessions/, archived_sessions/auth.json, config.toml, *.sqlite
OpenCodeopencode.db, read-onlyauth.json, opencode.json, the credentials tables
Antigravity CLItranscripts, history.jsonl, settings.jsonits MCP configuration and keychain entry
Pi~/.pi/agent/sessions/, models-store.jsonauth.json, models.json, settings.json

Full list in the README

Does it send anything anywhere?

No. No telemetry, no analytics, no calls of its own. Your CLI still talks to its own AI provider.

$ npx agent-workbench
----------------------------------------------------------------
  Agent Workbench
----------------------------------------------------------------
  URL             http://127.0.0.1:52341/?token=…
  Directory       D:\My App
  Available CLIs  claude-code, codex, opencode, antigravity, pi
A random port and a new token on every start.
  your browser
        |
        |  127.0.0.1, token
        v
+-------------------+
|  Agent Workbench  |
+-------------------+
        |
        |  a real terminal
        v
+-------------------+
|  your CLI         |
+-------------------+
        |
        |  its own login
        v
  its AI provider
The only line out is your CLI’s own.

What does my machine need?

Node.js and one CLI you already use.

node >= 20
22.13 or later for OpenCode’s history and Antigravity CLI’s titles.
claude | codex | opencode | agy | pi
At least one, installed and logged in. Agent Workbench finds it in your PATH.
git [optional]
For the Changes panel. Everything else works without it.
Windows, macOS, Linux
Used every day on Windows 11. On Linux, installing compiles the terminal module: python3, make and a C++ compiler.

How do I open a project?

Run one command from its folder. It opens in your browser.

Install with

npx agent-workbench

npm install -g agent-workbench

agent-workbench

git clone https://github.com/cvelasquez/agent-workbench.git

cd agent-workbench && corepack enable pnpm && pnpm install && pnpm dev

On Android 10 or later, the app opens the SSH tunnel and tells you when an agent finishes or waits for you.

The projects sidebar with sessions from several CLIs, each with its badge, and the + menu with the five CLIs and their versions