Experimental tool — active development

A controller for running coding-agent swarms.

Describe a project or paste in a task list. ShrimpHub queues the work, spawns coding agents, runs validation, and keeps enough state for long-running repair loops. It is an early tool for technical users who want visibility into every agent run.

Get started → Read the docs

ShrimpHub Dashboard

25k+

agents run

106

projects built

21.1B

tokens processed

348k

calls in the last 7 days

// field note

14 hours. 41 agents. One playable Godot game.

Chess-2 is a playable Godot 4 milestone built from a design document while the controller ran unattended. The useful part is that agents wrote code, repaired failures, accumulated project knowledge, and a final QA agent found a real gameplay bug.

Read the case study

41agent runs

14h 21munattended run

123passing GUT tests

69.8Mtokens consumed

// features

Write tasks.
Watch the queue.

Built around Godot, Python, and TypeScript projects. Primarily tested with MiniMax M2.7, with support for Claude, OpenRouter, Kimi, and custom OpenAI-compatible or Anthropic-compatible endpoints.

SHRIMP stands for Self-Healing Runtime for Iterative Multi-agent Programming — a system that keeps running even when individual agents fail.

// how to use it

Use Claude or Codex to steer. Use MiniMax agents to do the long run.

The current happy path is a MiniMax token subscription as the worker model. ShrimpHub was designed around high-volume agent runs: lots of implementation, QA, retries, and recovery work where cheaper model throughput matters.

Recommended MiniMax M2.7

Offload work from Claude or Codex

Keep Claude or Codex for planning, judgment, and review. Send scoped implementation, bugfix, QA, polish, and cleanup tasks to ShrimpHub, where MiniMax-backed agents work through the queue.

Hand off a whole sprint

Plan a sprint in Claude or Codex, convert it into a task graph, then let ShrimpHub run for hours or days. Come back to dashboard state, logs, diffs, completed tasks, and follow-up bugs.

Create projects inside ShrimpHub

Start from a project description, let the controller create the task graph, and run the project autonomously with MiniMax agents through implementation, validation, and repair loops.

Easy install path: ask Claude or Codex to install ShrimpHub from github.com/shrimplabs/shrimphub, configure MiniMax, and start the local dashboard.

// long-horizon planning

Most LLM tools are built for minutes of work.
ShrimpHub is built for longer runs.

State lives outside the context window, tasks chain across sessions, and agents can pick up repair work after failures. Longer runs still need supervision: logs matter, diffs matter, and some failures require human judgment.

ShrimpHub helped build ShrimpHub.

After the initial bootstrap, ShrimpHub's own agents wrote features, fixed bugs, and ran QA while the system was still changing. That made it useful quickly and gave us a detailed record of how the agents behaved.

View commit history →

Multi-agent orchestration

Spawns parallel LLM agents that work through a task backlog. Bounded concurrency helps control cost and blast radius, but you should still watch provider usage.

Self-healing chains

Failed agents retry up to 3× with failure context injected. If retries are exhausted a recovery agent picks up the thread automatically.

Dependency graph

Tasks chain automatically. Parallel work runs in parallel; sequential work waits its turn. Topological sort resolves the execution order.

File-level locking

Before an agent touches a file it acquires a lock. Other agents skip that file, eliminating merge conflicts across concurrent tasks.

Any LLM provider

Works with Claude, MiniMax, OpenRouter, Kimi, Ollama, or custom OpenAI-compatible and Anthropic-compatible endpoints. Switch providers in one config line.

Live observability

Full REST API on localhost:5001. Stream agent logs, inspect task state, and trigger spawns from the dashboard or your CI pipeline. The task graph updates dynamically as agents reason through post-task dependencies — new tasks can be added on the fly.

// how it works

Four steps.
Keep git close.

From plain-language description to queued work, agent runs, validation, and review. The controller does a lot; you still decide what to keep.

Describe your project

Type what you want built in plain language, or write tasks manually. ShrimpHub can turn a project brief into a task DAG.

Agents spawn

The orchestrator fills agent slots from the queue. Parallel tasks run in parallel — bounded by your max_active_agents setting.

Self-healing kicks in

Failed agents retry with failure context. If retries are exhausted a recovery agent picks up the thread. The system keeps running.

Review the result

The dashboard shows what ran, what changed, and what's left. Review the logs and diff before committing anything.

Agent Lifecycle

Works with every major LLM provider

Claude MiniMax OpenRouter Kimi Ollama Custom OpenAI-compat

Open source.
Run the controller yourself.

MIT licensed orchestration software for technical users. Bring your own inference provider, keep projects under version control, and follow the project as the APIs evolve.

Get started ⭐ Star on GitHub

⚠️

Early software notice. ShrimpHub runs autonomous LLM agents that can write code and run tools. APIs and workflows are still evolving. Primarily tested with MiniMax M2.7; Claude, OpenRouter, Kimi, and custom providers are supported. Use version control, keep important work backed up, and review agent output before relying on it.

language