Hi Product Hunt, Spore is my new Open Source project.

Spore is built around a few deliberate choices:

Write programs instead of paying the model to repeat a procedure. Agents write, test and keep helpers they can run on later tasks.
Keep the model-facing interface small. Seven built-in tools; everything else happens through ordinary programs, rather than an ever-growing catalogue of specialised tools.
Keep large outputs in files, not automatically in context. The agent searches and reads what it needs instead of carrying every command’s output into subsequent model calls.
Start with a general assistant that creates focused specialists. Delegation has two levels: the assistant and its cells. Cells cannot create more cells.
Enforce boundaries outside the agent. The kernel controls access, brokers hold credentials, and the runtime records spend and activity where the agent cannot rewrite them. Command-word blocklists and promises of good behaviour are not security boundaries.

An agent is a folder and a strictly checked definition file—not an application you have to implement. Its conversation, files and programs persist there. Anything that can write the definition can create an agent, including your application or another agent.

Those choices make Spore both a tool you can use directly and a runtime you can build into your software.
There are two ways to start:

Use it yourself: `spore here` opens a coding agent in a Git repository, with a terminal interface, persistent conversation, diff and undo.
Build on it: your application sends work through Node, HTTP or the CLI, follows events and collects results. Your application keeps its own interface and workflow.

The agent is untrusted; the operator is trusted. Linux kernel mechanisms enforce access boundaries, provider credentials stay outside the agent, and each turn has spend, time and model-call limits. Network access is open by default, so additional egress restrictions belong to the host.

To show the kind of work it can support, we published Blender experiments and a playable Godot game, with prompts, results, verification scripts and limitations:

https://rhizogrip.com/spore/demos

These are individual examples, not benchmarks or guarantees. The Godot work runs inside Spore’s sandbox; the Blender demo uses an application bridge that executes outside it.

Spore is Apache-2.0. Agents execute on Linux; macOS uses a Lima VM and Windows uses WSL2. You’ll need supported model access, and setup offers tested model recommendations.

I’d love to hear what you’d try first: an agent working in your own project, or one embedded in your application. Happy to help with getting started.

Source:

https://github.com/RHIZOGRIP/spore