Skip to content
PortBay
Comparison

PortBayvsSculptor

Sculptor and PortBay isolate agents the same way, which makes the rest of the comparison unusually clean. Sculptor's launch post in September 2025 argued for a container per agent and against git worktrees; the shipping product documented today does the opposite — its docs say "By default a workspace is a git worktree off your repo", with Clone and In-place as experimental alternatives and Docker reserved for an experimental mode that containerises Sculptor's own backend process. PortBay puts each dispatched card in a worktree too, at .portbay/worktrees/<card-id>. So the argument is not about isolation mechanism. It is about what is inside the isolated copy. In Sculptor, a checkout, a workspace-scoped terminal, a skills library and whatever your per-repo setup command installs. In PortBay, a checkout plus a running application: a pinned runtime, a database from seven engines, trusted HTTPS on a .test domain, captured mail — and a macOS Seatbelt profile with a network policy wrapped around the run.

PortBay vs Sculptor comparison page.
The short answer

Which one is right for you

Pick PortBay if

You want the full stack, open source.

The agent's work has to be provable against a live app. You want the worktree to contain a running project — a migration to apply, a page to load over HTTPS, an email to catch — with the runtime and the database provisioned rather than installed by a setup command you maintain. You want an operating-system boundary around the run, not just a separate directory: a generated macOS Seatbelt profile per project, with a network policy and logged denials. You want a kanban board whose cards are markdown files in the repo, an MCP server so the agent can update its own card, and the sites already configured in Herd, ServBay, MAMP or Valet imported instead of rebuilt.

Stick with Sculptor if

It already fits your workflow.

You want the agent workspace itself to be better, and you want it free. Sculptor is MIT-licensed with no paid tier at all, ships a curated skills library (spec, mocks and a /sculptor-workflow:fix-bug flow that runs a reproduction interview, writes failing tests, then makes them pass), runs several agents inside one workspace on a shared working copy so one can implement while another writes tests, and gives each workspace its own terminal, change review and GitHub pull-request tracking. It supports the Pi agentic harness alongside Claude Code and will run any terminal-based agent. It builds for Linux on both x86_64 and ARM64. If your bottleneck is agent workflow quality rather than the environment underneath it, Sculptor is doing more interesting work there than PortBay is.

Where PortBay and Sculptor actually differ

The table below has the row-by-row detail. These are the three shapes it cannot draw: the job you already use one of them for, the job only one of them does, and what survives the move.

Both run your sites

PortBay and Sculptor overlap on the everyday job: a local site on a real domain, over HTTPS, with the services it needs. If that is all you need, either will do it.

Only one dispatches agents

Move a card to To Do and PortBay launches Claude Code, Codex or Cursor against a project that is already running. Sculptor has no board to dispatch from.

Switching keeps your projects

Point PortBay at the folders you already have. They stay where they are, .test domains are re-issued on the first run, and databases are imported rather than rebuilt.

The detail

Feature by feature

Sculptor rows come from Sculptor’s own site and docs, on the date below. PortBay rows come from the app’s source. We mark partial support honestly — including where the other side wins.

Feature
PortBay
Sculptor
What each tool is
Primary job
Local dev env + agent board
Parallel agent workspaces
Runs a dev environment for the agent
Yes — the agent's wedge
Setup command + env vars
Sculptor's environment story is a per-repo setup command (e.g. `npm install`) run at workspace creation, plus env vars from ~/.sculptor/.env and .sculptor/.env. Useful, but it installs into a checkout — it does not start services.
Kanban board for agent work
Pending → Todo → Done
Workspaces + agent tabs
Sculptor's units are workspaces and agent tabs, not columns. PortBay's card moving to Todo is what triggers a run.
Product maturity
Shipping releases
Self-described preview
Sculptor's own README: "Sculptor is actively under development and should be treated as an experimental research preview… Things will not be perfect. Expect mistakes and bugs." Quoted, not characterised.
Platforms
macOS (ARM + Intel), Linux
mac ARM, Linux x64+ARM
Sculptor ships Mac (Apple Silicon), Linux x86_64 and Linux ARM64. No Windows and no Intel Mac build; the September 2025 post listed both as "on the way". PortBay ships Intel Mac but no Windows either.
Isolation & containment
Parallel agents in git worktrees
Yes — worktree per card
Yes — worktree by default
Same mechanism on both sides. Sculptor: "By default a workspace is a git worktree off your repo", on a <user>/<slug> branch. PortBay: .portbay/worktrees/<card-id> on a card/<id> branch, under a board-level concurrency cap.
OS-level sandbox around the run
Seatbelt + network policy
Worktree isolation only
A worktree isolates the code, not the process — an agent in a worktree still has your user's filesystem and network. PortBay adds a generated macOS Seatbelt profile per project: writes confined to the project folder, network policy set to blocked, loopback, outbound or full, denials logged line by line.
Docker option
No
Experimental backend only
Sculptor's Docker mode is experimental and containerises the Sculptor backend process — not one container per agent. It needs Python 3.10+ and Docker on the host, and the shipped recipe builds an image with git, the Claude CLI and runtime dependencies. PortBay has no Docker mode and does not need one.
Other isolation modes
Worktree or main tree
Worktree, clone, in-place
Sculptor offers Clone (a full separate clone with its own remotes) and In-place (the agent works directly in your repository) under Settings > Experimental.
Several agents on one working copy
One agent per worktree
Yes — agent tabs
A genuinely different design. Sculptor's agent tabs share a workspace: "All agents in a workspace share the same working copy", so one can build a feature while another writes tests for it. PortBay's cards each get their own worktree, which trades that collaboration for the guarantee that two runs can never collide.
Agent workflow
Agents supported
11 CLIs + any via Custom
Claude Code, Pi, any CLI
Sculptor integrates Claude Code and the Pi agentic harness, and will run any terminal-based agent. PortBay dispatches 11 clis + any via custom.
Bundled skills library
Your CLI's own skills
Yes — built-in workflows
Sculptor ships spec, mocks and /sculptor-workflow:fix-bug, which runs a short reproduction interview, writes failing tests, then makes them pass. PortBay ships no equivalent — it uses whatever skills your agent CLI already has.
Change review and PR tracking
Card comments + your diff
Review UI + GitHub PRs
Built-in terminal
No
Scoped to the workspace
Command palette
No
Cmd+K
MCP so agents update their own task
Yes — read/write cards
Not documented
The environment the agent works in
Managed PHP / Node runtimes
Yes, pinned per project
Setup command installs
Per-project database
7 engines, per project
No
PortBay provisions and supervises the engine — MySQL, MariaDB or PostgreSQL and four more — with an isolated data directory and injected connection variables. Sculptor documents no managed databases.
Trusted local HTTPS + real domain
mkcert on yourapp.test
No
Captured mail + public tunnel
Both, one click
No
Imports existing local sites
Herd, ServBay, MAMP, Valet
No
Licence, price & cadence
Price
Free to 6 projects; $10/mo
Free — no paid tier
Sculptor has no pricing page and no paid tier. PortBay's Community plan covers 6 projects; Pro is $10/month, or free for life with a merged qualifying pull request or a GitHub sponsorship.
Open source
Yes — AGPL-3.0
Yes — MIT
Sculptor is MIT, which is more permissive than PortBay's AGPL-3.0. Its README does add: "At this time we don't have the bandwidth to take on a large volume of external contributions."
Release cadence
Active
Active — v0.46.0
Checked 2026-09-06 against the GitHub API: latest release sculptor-v0.46.0 on 2026-09-01, last push 2026-09-04, repository created 2025-08-07.
Runs fully local
Yes — nothing leaves the Mac
Yes — desktop app
Both are local-first; the agent CLIs still talk to their own model providers.

Sculptor facts on this page were last checked against github.com/imbue-ai/sculptor on . Pricing and limits change — check their site before you decide.

Migration

Already using Sculptor?

Keep it, and put PortBay underneath it. The two do not collide: both dispatch agent CLIs already installed on your machine, both work in git worktrees off your repository, and neither claims your remote. Sculptor is where you shape how an agent works — the skills library, several agents on one working copy, the review pass before a pull request. PortBay is what the agent works against: point it at the same repository, press play, and the project comes up with a pinned runtime, a database, trusted HTTPS on a .test domain and captured mail, so a Sculptor agent editing that checkout has real services to hit. If you would rather run everything from one board, create the card in PortBay instead and let it dispatch the agent inside the Seatbelt profile.

Sculptor
1

Install PortBay and add the repository you have been opening Sculptor workspaces against.

2

Press play — PortBay provisions the runtime, a database and trusted HTTPS for that project.

3

Either point a Sculptor workspace at the same repo and let its agents use the live services, or create a PortBay card, assign Claude Code, and move it to Todo to run inside the sandbox.

PortBay
Common questions

PortBay vs Sculptor, in plain terms

Don’t see your question answered here? The docs go deeper — every feature, flag and failure mode.

Read the docs

Not by default, and not any more in the way its launch post described. Sculptor's September 2025 relaunch argued for a container per agent and specifically against git worktrees. The shipping product documented on 2026-09-06 says the opposite: "By default a workspace is a git worktree off your repo", with Clone and In-place as experimental alternatives. Docker still appears, but as a separate experimental feature — "Running the Backend in a Container" — where Sculptor's own backend process runs inside Docker, over SSH on a remote, or in a VM, via a custom backend command that needs Python 3.10+ and Docker on the host. That is not a container per agent, so it is not the isolation boundary the older post described.

Both put the agent's code in a git worktree, which isolates the checkout but not the process — an agent running in a worktree still has your user's filesystem and network access. PortBay adds the operating-system boundary: it generates a macOS Seatbelt profile per project that confines writes to the project folder and applies a network policy of blocked, loopback, outbound or full, logging every denial. Sculptor documents no OS-level sandbox for the agent itself. If your worry is a bad command reaching outside the repository, that is the difference; if your worry is only losing your uncommitted work, a worktree already handles it and both tools are equivalent.

Several things, and they are the reasons to pick it. It is free with no paid tier and MIT-licensed. It runs multiple agents inside one workspace on a shared working copy, so one can implement a feature while another writes tests for it — PortBay gives each card its own worktree, which prevents collisions but also prevents that collaboration. It ships a curated skills library including a fix-bug workflow that interviews you for a reproduction, writes failing tests, then makes them pass. Each workspace gets its own terminal, a change-review pass and GitHub pull-request tracking, and there is a Cmd+K command palette. It supports the Pi agentic harness alongside Claude Code, and it builds for Linux on ARM64 as well as x86_64.

PortBay runs the environment. A PortBay card dispatches the agent against a project that is already up: a pinned PHP or Node runtime, a database from seven engines with its own data directory and injected connection variables, trusted HTTPS on a real .test domain via mkcert, captured outbound mail and a one-click Cloudflare tunnel. It confines that run in a macOS Seatbelt profile with a per-project network policy. It gives you a kanban board whose cards are markdown files under .portbay/tasks/, and an MCP server so the agent can update its own card while it works. And it imports the sites you already have configured in Laravel Herd, ServBay, MAMP or Laravel Valet. Sculptor's environment layer is a setup command and an env file.

Yes. There is no pricing page and no paid tier; the repository at github.com/imbue-ai/sculptor is MIT-licensed, which is more permissive than PortBay's AGPL-3.0. Two caveats worth knowing, both in Imbue's own README: Sculptor is described as "an experimental research preview" where you should "expect mistakes and bugs", and Imbue says it does not currently have the bandwidth to take on a large volume of external contributions. PortBay is free on the Community plan for up to 6 projects, $10/month for Pro, or free for life with a merged qualifying pull request or a GitHub sponsorship.

PortBay, by count, though both are open about it. Sculptor integrates Claude Code and the Pi agentic harness and will run any terminal-based agent you configure. PortBay dispatches 11 out of the box — Claude Code, OpenAI Codex, Cursor, Gemini, Aider, Copilot CLI, OpenCode, Amp, Qwen Code, Antigravity and Ollama — plus any other CLI via a Custom command template. In practice this matters less than the raw numbers suggest — if you only ever run Claude Code, both detect it and use the subscription you already pay for.

Yes, and it composes well. They store state in different places, both work in worktrees off your repo, and neither takes ownership of your git remote. The natural split: PortBay brings the project up — runtime, database, .test domain, mail catcher — and Sculptor's agents work against those live services from their own workspaces. Nothing about running one prevents the other, and neither changes your agent CLI setup or your CLAUDE.md.

PortBay mascot — a friendly blue tugboat

Give your projects and your agents a real local home.

Download for macOS

Free & open source · macOS 11+ on Apple Silicon · Pro from $10/mo