Sculptor
A desktop app for running coding agents in parallel across isolated workspaces, then reviewing and merging their changes.
- Source repo
- imbue-ai/sculptor
- Stars
- ★ 244
- Last updated
- today
- License
- MIT
- Primary language
- Python
- FA score
- 44/100 · Major gaps
At a glance
- Works with
- Portable with changesClaude Code
- You'll need
- Typical use
- A full-stack engineer who wants spec writing, mocks and bug fixes running at once across separate parallel workspaces.
- Main limitation
- Officially labeled an experimental research preview: expect bugs and significant, fast-moving changes.
- Source review
- 44/100 · Major gaps
What does this agent do, and when should you use it?
Sculptor is a desktop application from Imbue for running coding agents in parallel. You connect a repository, create a workspace (an isolated copy of the code with its own branch and setup), and prompt an agent; when it finishes you review the changes and merge back to main. It ships with bundled skills that structure agentic workflows, plus a built-in workspace-scoped terminal, a Cmd+K command palette, a Changes review surface, and GitHub pull request creation and tracking. Sculptor has integrated support for the Pi agentic harness in addition to Claude Code, and it can also run and manage any terminal-based agent. It is MIT-licensed, maintained by Imbue, and explicitly framed as an experimental research preview that will change quickly.
Sculptor connects to a Git repository and creates workspaces — isolated copies/worktrees of the repo with their own branches and setup. Inside a workspace, the Chat panel takes a prompt along with model, context, plan/effort and mention settings, and hands the task to an agent such as Claude Code, the integrated Pi harness, or any terminal-based agent. The agent works through the workspace's built-in terminal, running commands and editing files. When the run ends, the user reviews the diff in Changes, commits, and can open and track a GitHub PR from Pull Requests. Bundled skills provide command-driven workflows, for example /sculptor-workflow:fix-bug, which runs a short reproduction interview, writes failing tests, then makes them pass. An experimental Container Backend can run the workspaces in Docker or on a remote host.
- A full-stack engineer who wants spec writing, mocks and bug fixes running at once across separate parallel workspaces.
- A developer fixing a UI bug end to end with
/sculptor-workflow:fix-bug, from reproduction interview to failing tests to a passing fix. - An existing Claude Code user who wants a graphical workspace manager and change-review surface around the agent.
- A team that prefers the Pi agentic harness, or wants to plug its own terminal-based agents into one managed UI.
- Someone experimenting with Docker or remote execution to isolate agent runs from the host machine.
- A small team that wants agent-generated changes to flow through a disciplined review, commit and GitHub PR path back to main.
How do you install or deploy this agent?
Download the build for your platform from the official links:
# Mac (Apple Silicon)
https://tryimbue.link/sculptor-for-apple-silicon
# Linux
https://tryimbue.link/sculptor-for-linux
# Linux ARM64
https://tryimbue.link/sculptor-for-linux-arm64The README provides these download entries but does not document a package-manager or source-build install command. For a local source setup it points to docs/development/getting_started.md; that file's contents are not part of this source, so the exact toolchain versions and build commands cannot be confirmed here.
How do you use this agent?
- Launch Sculptor and connect a local Git repository.
- Create a workspace — an isolated copy of the repo with its own branch and setup.
- Prompt the agent in Chat, choosing the model, context and agent harness (Claude Code, Pi, or any terminal-based agent).
- Let the agent work through the workspace's built-in terminal, editing files as it goes.
- Review and commit in Changes, then merge back to main; open a PR and track it under Pull Requests if you use GitHub.
- Start another workspace for a parallel task on the same project, or add another agent to an existing workspace to collaborate.
Prompt → agent run → Changes review → commit / PR → merge to mainExperimental remote execution requires enabling the Container Backend as described in docs/help/experimental/container_backend.md (Docker or a remote host).
What are this agent's strengths and limitations?
- Isolated-worktree parallelism is the core execution model, so multiple agents can work the same repo without stepping on each other.
- Supports both the integrated Pi agentic harness and Claude Code, and can additionally drive any terminal-based agent, avoiding a single-harness lock-in.
- Bundled skills encode real workflows, notably
/sculptor-workflow:fix-bugwith its reproduction-interview and failing-test-first loop. - Closes the loop from changreview to commit to GitHub PR instead of stopping at a chat pane.
- MIT-licensed and used extensively by Imbue itself.
- Prebuilt installers for macOS Apple Silicon and Linux, including Linux ARM64.
- Officially labeled an experimental research preview: expect bugs and significant, fast-moving changes.
- Maintainers state they lack bandwidth for a large volume of external contributions, limiting community involvement.
- No Windows download is listed; only macOS Apple Silicon, Linux and Linux ARM64 are offered.
- Model/provider details beyond Claude Code and Pi are not specified, so you must bring and pay for your own model access.
- Source-build instructions are deferred to a separate development doc not present in the repository, so the local build path is unverified here.
- The Container Backend for Docker or remote execution is explicitly experimental.
How does this agent compare with similar options?
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| Sculptor This agent | 44 · Major gaps | — | ★ 244 | today | Python | Claude Code |
| PandaCode | 58 · Major gaps | — | ★ 1.1k | today | Rust | Codex · Claude Code · Claude.ai |
| Poracode | 55 · Major gaps | Desktop appFree + model costs | ★ 121 | today | TypeScript | ChatGPT · Claude Code |
| dmux | 69 · Some gaps | CLIFree + model costs | ★ 1.8k | 2mo ago | HTML | Codex · Claude Code · OpenAI API |
How does FollowAgents rate this agent?
Why each dimension lost points
SECURITY.md honestly discloses that agents act with the user's access, opening a repo runs its code, plugins run with full UI privileges, and the loopback port is the trust boundary, recommending the container backend for isolation. That is candid disclosure, not control: no sandbox by default, no per-action confirmation, no rollback mechanism, so least_privilege/user_confirmation/rollback score 1. Data-flow transparency has a Privacy settings note and threat model but no concrete data-destination inventory, scoring 2. Dependency security shows only pyproject and workflow hardening rules with no lockfile audit evidence, scoring 1. Source attribution is a bare Imbue mention with no version/commit provenance, scoring 1.
Docs and the fake_claude tests show defensive parsing of missing fields and error handling, giving self_consistency 2; dependency availability lists Homebrew/tmux/just/uv/nvm prerequisites but no version pinning or recovery guidance, scoring 1; failure messages appear as is_error flags in tests but user-facing error copy is thin, scoring 1.
README clearly frames the audience (parallel coding-agent users) and scenarios (workspaces, PRs, skills), and capability_boundaries are set by the experimental-preview and container-backend notes, scoring 2; trigger precision (when a skill or agent fires) lacks explicit rules, scoring 1; environment fit covers only Mac/Linux with limited Windows or remote deployment detail, scoring 1.
Information architecture is clear with a docs index and concrete install steps, scoring 2; naming stability is explicitly unstable (drastic changes possible) and examples/FAQ are limited, each scoring 1; known limitations are well disclosed, scoring 2; MIT license is complete, scoring 2; no version number or changelog, scoring 0; maintenance responsibility names Imbue but limits external contributions, scoring 1.
Output usability (workspaces, change review, PRs) is described concretely, scoring 2; marginal value over bare agent use (parallel workspaces, skill library) is explained, scoring 2; cost-benefit mentions token budget but offers no quantified evidence, scoring 1.
README claims are mostly product description without traceable benchmarks or data, so claim_traceability scores 1; cross-source corroboration is limited to docs citing each other, scoring 1; fact/inference separation is decent in SECURITY.md but product claims are not so separated, scoring 1.
- No sandbox by default: agents read/write files and run commands with the user's access, and opening a repo executes its code; stronger isolation exists only in the experimental container backend.
- Plugins run with the same privileges as the UI, and remote plugin sources are re-fetched on every load, creating supply-chain risk.
- The loopback API is reachable by any process on the same machine; do not run on shared hosts or with exposed ports.
- The project self-describes as an experimental research preview subject to drastic change, with no version number or changelog, raising upgrade risk.
- Publisher identity is unverified by the FollowAgents registry, leaving maintenance responsibility and update path unclear.