Hi, over the past 6+ months I have built an ecosystem (libraries, CLI, Desktop) of tools to build or use AI agents.

Depending on how close or not close you are with coding, you may lean towards the libraries (compilr-dev/agents and compilr-dev/agents-coding) to build your own AI Agents; you may use the CLI to manage your project and run your agents in the terminal when working, for example in VS Code or you may use the Desktop app (it shares the same core of the CLI but has for sure a richer UI).

What, I believe, sets the compilr-dev ecosystem apart from other similar tools is the real multi-agent nature.

Focusing on the Desktop app, it ships with pre-built specialized agents (built using the above mentioned libraries) each with tailored system prompt, skills and access to dedicated tools.

When a new project is created a team of specialized agents is added to the team roster and you can add or remove agents already at project creation time (you can even create custom agents at project creation time).

You can choose your preferred LLM provider (BYOK approach) and you can seamlessly switch provider whenever you want.

Each agent has its own context but agents within a project also share some context. They can interact with each other by inquiring (agent A can ask something to agent B) or by task handoff.

Each agent is instructed to follow a spec driven approach and the app itself encourage the user to design before building. For example, as soon as you create a new project you can use the /design command to have the agent ask you questions to capture the project scope and details. As a result the agent will track work items in the database and will write the PRD document following your direction.

All files are visible and editable by you and by the agents.

Last but not least, special attention was given to the token consumptions: tokens = money so saving them is "paramount" (no, AI did not write this sentence, I deliberately chose to use the word "paramount"). Some techniques: - Tools and skills definitions are loaded only when needed (like Neo loading the kung fu skills on Matrix) - Tools output are pruned/summarized in the context after a few rounds - Agents spawn subagents to read large files and get a summary - and more

I just launched the compilr-dev ecosystem in beta and I hope to start having some beta testers to get some real world feedback.

Thank you.

Carmelo

Show HN: Compilr.dev, multi LLM AI workspace