OpenJellyfish
Readable intelligence.
Quick Start
Three ways, pick what fits you
Download & Run · One-Click Install
Desktop app, ready to use, perfect for individual exploration
M1/M2/M3/M4 chips
Intel-based Mac
x64 installer
Portable, reproducible, interpretable intelligence.
Instead of piling more onto an ever-heavier Agent, we split it into four layers: the LLM drives the intelligence, the Harness gives it a body to run in, Documents keep it unique at the core, and the Environment projects how it shows up.
Runtime + Prompt + Memory + Workflow = AgentLLM + Harness + Documents + Environment = AgentLLM
The DriveCan it think?
Provides reasoning power. Reasoning, language and world knowledge all come from here — swappable at any time: GPT, Claude, local models, plug-and-play via OpenAI compatibility.
Harness
The BodyHow it thinks
Decides the runtime mechanics. It only runs the Observe · Think · Decide · Act loop, with no idea who it is or what to do. It stays simple forever.
Documents
The CoreWhat it thinks, and why
Defines the intelligence itself. Prompt, Memory, Skills, Workflow, Permissions, Knowledge — all documents: readable, editable, copyable, version-controllable.
Environment
The StageHow it reaches the world
Decides how the intelligence interacts with the outside world. Filesystem, Tools, MCP, Scheduler, External Services — the intelligence reaches the world and shows its results through it.
The biggest difference from traditional Agent frameworks:Documents are not a config file for the Harness — they are the intelligence itself.
Portable Intelligence
Because the Agent is documents, copying a directory copies the whole Agent. Its entire lifecycle is managed with mature software-engineering tools:
The Harness can upgrade, the LLM can swap, the runtime can be rebuilt — but the soul lives on in the documents.
git clone→copy the whole Intelligencegit diff→see how the Agent grewgit merge→fuse two AgentsDocument is all you need
Every capability comes down to one thing — coworking on documents with the agent.
Paths Are Boundaries
Isolate how agents are built and shared by document path. Permissions and power boundaries come for free — a different boundary is a different agent.
Memory Palace
Memory at every level lives as documents — browse, search and edit it like walking through a memory palace.
Docs as Ops
The agent schedules cron jobs and runs system operations by writing documents — ops is just editing files.
Self-Rewriting
Documents are editable, and documents are the agent — so the agent can rewrite its own prompts, scripts and logic.
Experience → Agent
Turn human experience into a publishable agent fast — all you need is to cowork on documents with the agent itself.
Copy & Import
A directory is a whole agent. Copy, import and share it with a plain cp or git clone.
… and ?
Documents can do far more. The next use case is yours to define.
One step back. A little simpler.
In an ever-more-complex agent ecosystem, the jellyfish chooses to step back — but back into a closed loop.
OpenClaw
Hyper-private by design — high risk of agent overreach, no real multi-user story, and distribution is out of the question.
Hermes
Full automation sounds great, but you can only guide it through conversation — config, memory, and distribution silos are still yours to stitch.
OpenJellyfish
Built for tight, high-quality ops (≤50 people). Like Cursor's co-pilot, your agent sits beside you helping you build the next agent.
Flowise
Node-based prototyping. Maintenance gets hard once logic grows.
n8n
A workflow automation giant — AI is just one of its later-added nodes.
Dify
Building workflows / chatflows is painful. The graph structure is complex, and the best use case is still unclear.
Light, yet closed-loop. The jellyfish answer to 'simpler.'