Imbue Opens Sculptor, a Free App Running Coding Agents in Parallel

Imbue has open-sourced Sculptor under MIT. The desktop app runs Claude Code, Pi and other coding agents in parallel, so your workflow isn't tied to one AI lab.

·
·
·
Imbue Opens Sculptor, a Free App Running Coding Agents in Parallel
Read6 min
TypeNews
SubtopicCode Agents
  • Imbue open-sourced Sculptor, a desktop app for running coding agents in parallel, under the MIT license.
  • It works with Claude Code and the open-source Pi harness, and can run any other terminal-based agent.
  • It's free and runs locally. You bring your own Claude plan or open model, and Imbue doesn't train on your code.
  • Deep harness integration lets agent turns survive crashes and adds native plan panels and a context-usage chip.
  • Two earlier prototypes were scrapped as stronger models made their extra scaffolding pointless, as the launch post explains.
  • It's still a research preview: Mac and Linux only, with limited capacity for outside contributions.

Imbue has released the source code for Sculptor, a desktop app for running several coding agents in parallel. Developed internally for roughly two years, it runs on a developer’s machine, supports multiple agent harnesses and model providers, and carries an MIT license. The app is free, while model subscriptions and API usage remain separate costs.

Sculptor handles the operational work around parallel agents: creating isolated checkouts, tracking each session, syncing selected changes into an editor, and presenting diffs for review. Imbue also published a design retrospective that documents two abandoned prototypes and explains why the current product emphasizes parallel execution, local control, and model portability.

Discarded prototypes set the roadmap

The first prototype, Crafty, was built when GPT-4 represented the frontier of coding models. Imbue added extensive scaffolding to compensate for the model’s weaknesses, including tools that detected and repaired linter errors. Subsequent models handled many of those tasks directly, making the surrounding code redundant. The experience pushed the team toward infrastructure that would remain useful as models improved.

The second prototype, Blackberry, asked developers to write a nested plan of commit messages before agents implemented each commit. Imbue abandoned the approach after models became capable of drafting and revising their own plans. The workflow also conflicted with how developers usually create commits and omitted the chat interface that had become standard in AI coding tools.

Two design choices survived those experiments:

  • Parallel execution: Developers can start another task while an agent continues working, with Sculptor managing the isolated checkouts.
  • Pairing mode: Selecting an active task brings its changes into the developer’s checkout and preserves the agent session for continued collaboration.

Parallel runs, one review loop

  1. Create a workspace by selecting a repository, agent, and base branch.
  2. Start one or more tasks, each operating in an isolated Git worktree.
  3. Continue other work while Sculptor manages the worktrees and agent sessions.
  4. Open pairing mode to inspect or edit a selected agent’s changes in the local editor.
  5. Review the diff, resolve any issues, and merge the approved work.

The default interface includes code, terminal, agent list, chat, and plugin panels. The repository also provides reusable workflow skills. For example, /sculptor-workflow:fix-bug asks questions to establish a reproduction, writes failing tests, and then modifies the code until those tests pass.

Current workspaces use Git worktrees, which provide separate working directories connected to the same repository. Earlier versions placed every agent inside its own Docker container, giving each task an isolated runtime, dependencies, and ports. Docker-based execution remains available as an experimental backend.

Harness hooks add deeper control

A coding-agent harness is the program that sends prompts to a model, gives it tools, and manages the action loop. Sculptor integrates with that layer instead of treating the agent as an opaque terminal process. Its harness documentation says Claude Code runs as a streaming JSON process with its control protocol enabled, allowing Sculptor to receive structured events and send instructions during a session.

What the control channel enables

  • Session recovery: An interrupted turn can survive an app exit or crash.
  • Native interaction panels: Sculptor replaces Claude Code’s question and planning tools with versions rendered inside the app.
  • Dependency graphs: Agents can record task dependencies, which Sculptor displays as a graph.
  • Context tracking: A percentage indicator shows how much of Claude Code’s context window is in use after each turn.

Sculptor also integrates with the open-source Pi agent, which connects to several model providers. Other terminal-based agents can run through a generic interface, though they receive fewer app-level controls because Sculptor cannot read their structured events or replace their internal tools.

Capability Claude Code Pi
Native question and plan panels Yes, through replacement tools Yes, through an extension
Mid-session model switching Applied on the next message Applied between turns
Fast mode Yes No
Context usage indicator Yes No

Portability drives the design

Imbue built Sculptor as a vendor-neutral control layer so workflows, skills, task history, and interface choices can remain stable when developers change models. That architecture becomes more useful as open-weight coding models, whose parameters can often be self-hosted, improve enough to compete on routine development tasks.

Model changes still depend on the selected harness and provider. Claude Code applies a switch on the next message, while Pi can change models between turns. Imbue’s community wishlist also proposes an agent that would use a Claude Max allowance until its limit is reached and then continue through Pi with an open model. That automatic fallback is a proposed project, not a built-in feature.

Costs, platforms, and limits

License MIT
App price Free
Model costs Developer-supplied subscription, API access, or compatible self-hosted model
Supported builds Apple Silicon macOS, Linux, and Linux ARM64
Windows No build available
Release status Experimental research preview

Imbue says it does not store repositories or train on users’ code. The app and workspaces run locally, but cloud-hosted models may still receive prompts, source code, and tool output. Data handling therefore depends on the chosen model provider; fully local processing requires a compatible locally hosted model.

Preview warnings remain

  • Breaking changes are likely: The README describes Sculptor as an actively developed research preview and warns about bugs and substantial revisions.
  • External contributions are constrained: Imbue says the team currently lacks capacity to review many outside pull requests.
  • The repository changed significantly: Earlier reviews found a public repository containing only documentation and issue tracking. It now contains the application source and more than 2,100 commits.
  • The community is small: At publication, the repository had roughly 246 stars and 20 forks.

Who gets value today

Sculptor best fits developers who already run several coding agents and spend time creating worktrees, monitoring terminals, and moving changes between branches. It also provides a consistent environment for comparing Claude Code with open-weight or alternative hosted models on the same repository.

Teams can audit and modify the application code, and Imbue regularly assigns Sculptor agents to improve Sculptor itself. That practice can produce customized versions shaped around a team’s preferred workflows, tools, and review conventions.

Developers who need stable interfaces, broad platform support, responsive support, or a large plugin ecosystem may find the preview limitations restrictive. Claude Code, Codex, and Cursor currently offer more mature product paths, while Sculptor offers source access, parallel-agent orchestration, and greater control over model selection.

Trending
  • No trending articles

Comments

avatar

Next Reads