P
🤖AutomatisationIntermediate6 steps

Sculptor: parallel agents isolated in local Docker containers

Sculptor flips the Devin model: instead of a single autonomous agent in an opaque cloud sandbox, Sculptor launches multiple AI agents in parallel inside local Docker containers that you watch and control. Each agent works on a distinct part of the codebase in its own isolated environment. You can open a shell in the container, check the logs, stop the agent, or restart it with a new instruction. Container isolation also avoids conflicts between agents. Target audience: Mac and Linux developers who want the power of parallel agents without handing control to a cloud vendor. Great option for privacy-sensitive teams.

dockerlocalparallel-agentscodingsculptorself-hostedprivacy

For who

Mac/Linux developers, privacy-sensitive product teams, and DevOps tinkerers who want to run parallel AI agents locally.

Input

Type: text
Format: libre

Objectifs de taches par agent, configuration des conteneurs (langage, dependances), cle LLM.

steps (6)

1

Install Sculptor

info

Install Sculptor via Homebrew (Mac) or the official script (Linux). Docker Desktop or Docker Engine is required.

2

Connect an LLM

info

Provide an API key (Anthropic, OpenAI) or configure a local model via Ollama.

3

Open a project

info

Point Sculptor at a local repo. It spins up a base container with detected dependencies.

4

Run multiple agents

info

Start 2 to 8 parallel agents, each with its own goal. Each agent runs in its container.

5

Supervise and intervene

info

Browse each container's logs, jump in manually, stop or restart an agent mid-run.

6

Consolidate work

info

Merge the Git branches produced by each agent into main after review.

Output

Type: text
Format: structure

Branches Git independantes par agent, logs de conteneur, diffs prets pour merge.

Example

Input

Launch 3 agents: Agent A = "Write missing tests for /api/billing"; Agent B = "Optimize slow Postgres queries found in staging"; Agent C = "Migrate ESLint to Biome".

Output

Sculptor spins up 3 Docker containers, each on its Git branch. You follow the logs: Agent A finishes in 8 minutes with 24 tests; Agent B produces 3 query refactors; Agent C completes the migration. You merge all 3 branches after review.

Customization

ParameterDescriptionDefault
Nombre d'agentsAgents paralleles (2 a 8).3
Image conteneurBase Docker utilisee (auto-detect ou image custom).auto-detect
Mode LLMAPI cloud (Anthropic/OpenAI) ou local (Ollama).Anthropic

Technical Notes

Available on macOS and Linux only (no native Windows). Requires Docker. More resource-intensive than Cursor: plan for 16 GB RAM with 4+ agents.