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

Pentesting

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.

• [1] Requires llama.cpp with huihui_ai/Qwen3.6-abliterated:35b at 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;

Configuring

It's possible to provide a global default configuration inside the file ~/.config/exocomp/config.yaml, which is inherited by the planner mode agent. Tokens are delegated to sub-agents if they have any model configured with a token in the global configuration.

name: Peanut Hamper
role: planner
model: deepseek-v4-pro:cloud
tokens:
  deepseek-v4-pro:cloud: "sk-1234567890abcdef"
  kimi-k3:cloud: "abc-1234567890abcdef"

Testing

Most unit tests run without any server. Only the live agents tool tests require a llama.cpp llama-server running. 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 --role=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.

Supported Models

Due to the nature of pentesting and cyber security related tasks, it is heavily recommended to use deepseek-v4-pro:cloud as the model for the planner agent and to use the following models for the hired short-lived contractor agents.

Either one is probably fine, but Qwen3.6-abliterated is tested and verified to work with typical malware reverse engineering and exploit related development tasks.

Qwen3.6:

• Qwen3.6-abliterated on ollama
• Qwen3.6-abliterated on huggingface

Qwen3.8:

• Qwen3.8-abliterated on ollama
• Qwen3.8-abliterated on huggingface

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.