The universal agent development environment

Run a team of coding agents. Keep the judgment.

iCode Workbench is the universal ADE: every agent, every platform, one workbench. Run Claude Code, Codex, and OpenCode in parallel, each in its own Git worktree. Mira, your cross-project assistant, watches every repository, tells you what needs you, and can act across them. Review the diff, and keep the terminal alive when the app restarts.

Local-first · Pre-1.0 · iCode Agent available today

The iCode Workbench All projects wall showing live sessions, tasks that need attention, and last activity across three projects
The Projects wall: in flight, needs you, and last activity across every repository.

Two surfaces

One environment for the whole loop.

Workbench is the desktop control room for a team of agents. Agent is the focused terminal surface for one task at a time. Both work in your real repository, and both leave the decisions to you.

Desktop · ADE · Early access

iCode Workbench

Run agents in parallel, queue durable tasks, review changes in the session worktree, and let Mira bring the blockers to you. A detached host keeps terminals alive across app restarts.

  • Claude Code, Codex, and OpenCode — each in its own worktree
  • Tasks, approvals, schedules, and task history
  • Source Control and an editor in the same checkout
Explore iCode Workbench →
Terminal · CLI

iCode Agent

A focused coding agent that runs the explore → change → verify loop directly in your repository. Context is defended at every boundary so long sessions stay coherent and affordable.

  • Seven tools: read, write, edit, grep, find, ls, bash
  • Print mode and JSON event streams for scripts and CI
  • OpenAI- and Anthropic-compatible model endpoints

npm install -g @icodehq/icode-agent

Explore iCode Agent →

Meet Mira

A teammate who watches all your projects.

Mira is a person in the rail, not another project tab. One conversation spans every repository: she reads the real queue, answers “what needs me?”, and can create, cancel, retry, or archive work through validated actions.

She has no shell and no project working directory. She cannot approve a proposal, accept a review, merge, or push. Those stay human decisions.

  • She comes to you — approvals, failures, waiting tasks, and review-ready work arrive with a rail badge and an OS notification when no window is focused.
  • She acts with limits — low-risk tasks can auto-dispatch under a daily cap; protected paths and keywords wait for a person.
  • She remembers — durable preferences live in Settings, visible and removable.
Mira streaming a Markdown status report with a task table, a live thinking-token count, and a Stop button
Mira streams a cross-project status report while the work line keeps the harness stage visible.

Principles

The work stays yours.

Isolation

Agents do not share a checkout.

Every new session gets its own Git worktree and branch at the current HEAD. Two agents in the same repository never edit the same files by accident.

Judgment

Review is a gate, not a formality.

Task work shows its changes against a baseline, and a review decision is separate from process exit. Accepting records the judgment; it does not merge, push, or delete anything.

Local-first

Your machine holds the state.

Projects, sessions, worktrees, tasks, and schedules live in local SQLite. There is no telemetry and no hosted service; model traffic belongs to the harness you chose.

Start with the surface that fits.

Use Workbench when you want a team and a control room. Use Agent when you want one focused session in the terminal. Neither takes the judgment away from you.

iCode · The agent development environment.