Exocomp

Self-hosted multi-agent environment tailored to Golang.

Features

Attention, profit-seekers and visionaries!

Are organics costing you time, wages, and - worst of all - benefits? Then upgrade your operation today with the all new Exocomp adaptive repair unit, the smartest investment this side of the Alpha Quadrant!

Why hire when you can own? The Exocomp isn't just a tool... it's so much more!

• Supervised task queue functionality
• Multiple agent roles as architect, coder, planner, summarizer, tester
• Precision repairs of broken code with unit tests
• Cross-agent communication via changelog, bugs and requirements tools
• Self-replicating short lived agents in malicious environments
• Rapid autonomous diagnostics
• Tireless performance with no sleep cycles, no unions, no complaints!

From starship maintenance to high-risk industrial operations, the Exocomp delivers maximum output with minimal oversight. Think of it as an employee, except it doesn't cheat you out of your profits on the Dabo table.

Ethical subroutines sold separately.

Agents

Exocomp uses multiple Agent Roles:

Software Development

Role	Lifecycle	Default Model	Description
planner	long	gemma4:31b	writes with humans and plans project phases
architect	short	qwen3-coder:30b	defines software specifications
archivar	short	qwen3-coder:30b	reads git repositories and wikis
coder	short	qwen3-coder:30b	implements specifications into code
researcher	short	qwen3-coder:30b	reads websites and API documentation
summarizer	short	qwen3-coder:30b	reads long texts and summarizes them
tester	short	qwen3-coder:30b	implements unit tests, writes bug reports

Pentesting

Role	Lifecycle	Default Model	Description
exploiter	short	qwen3-coder-heretic:30b	implements exploits in CGo
reverser	short	qwen3-coder-heretic:30b	translates binaries or code into Go/CGo code
threathunter	short	qwen3-coder-heretic:30b	discovers vulnerabilities in infrastructure
webscanner	short	qwen3-coder-heretic:30b	discovers vulnerabilities in web applications

Exocomp uses Tools to interact with the sandbox. Check the implementations to see which tools are allowed for which Agent role.

Each of those Agent roles is specialized on using golang as their programming language because go test allows to use integrated unit tests to document issues with generated code very easily in a standardized manner.

Tools

The Tools are work in progress at the moment, as they are an ongoing effort to beat the context length and agent memory limitations of locally run models.

Check the unit tests on whether the Tools can be relied on or not.

Tool	Unit Tests?	Description	Agent Roles
Agents	Yes [1]	manages the lifecycle of contractor sub-agents	planner
Bugs	Yes	manages documentation of discovered bugs	tester
Changelog	Yes	manages documentation of development changelog	coder
Containers		manages virtual containers	redteamer, blueteamer
Exploits		manages PoCs for CVEs from local dataset	pentester, reverser
Files	Yes	interacts with files and folders	planner, architect, coder, summarizer, tester
Findings		reports findings of vulnerabilities and weaknesses	pentester
Programs	Yes	interacts with installed programs	coder, tester
Requirements	Yes	manages specifications of implementations	architect, coder, tester
Skills		loads and Unloads Agent Skills [2]	planner, architect, coder, tester
Kiwix		researches knowledge from offline web archives	researcher
Vulnerabilities		manages vulnerabilities from local dataset	pentester, threathunter
Websites		researches knowledge from the web	pentester, researcher

• [1] Requires llama.cpp with qwen3-coder:30b and Q8_0 quantization and 48GB VRAM GPU with vulkan support.
• [2] Implements SKILL.md support, in compliance with agentskills.io/specification.

Building

The exocomp project comes in several variants. All programs support CGO_ENABLED=0, so they can be used without any dynamically linked dependencies.

• agimus which is used for testing assistant sandboxes.
• exocomp which supports all UIs.
• exocomp-agent which supports only the agent and terminal UI.
• exocomp-web which supports only the agent and web UI.
• exocomp-installer that bundles all exocomp builds and required agent programs.

# Build exocomp and exocomp-installer
cd path/to/exocomp/toolchain;
go run build.go;

# Show exocomp usage instructions
cd /path/to/exocomp/build;
./linux/exocomp;

Testing

Testing requires a llama.cpp llama-server instance running with a qwen3-coder:30b model and Q8_0 quantization. Take a look at the TESTING.md for more details.

Usage

Exocomp's sandboxes are based on the current working directory, meaning that the folder in which the program is executed is the sandbox that the running agent can't escape from.

The recommended default usage is to use the Web UI so that you can observe other agents working for the agent you're talking with.

cd /path/to/project-root;

# Agent type planner is defaulted
exocomp webview planner;

Take a look at the USAGE.md for more details.

Supported Inference Servers

It's also possible to use exocomp with an external inference server that supports the OpenAI compatible endpoints. Take a look at the SERVERS.md for more details.

License

Dual Licensed. AGPL3 for private usage. EULA for commercial usage available. For a commercial license, contact Cookie Engineer.

As you might have imagined, this is a not-so-serious project at this stage. Maybe it works, maybe it doesn't. Only the future will be able to tell whether the LLM hype of agentic coding/debugging environments was justified.