You can try it for free without signing up. If you create an account, it creates a track record of your activity, similar to the Github heatmap. Once you have enough activity, it calculates badges for consistency and AI skills (more coming).
I personally use it a lot through the mcp server so I can have my coding agent create and update tasks as it's working. It then becomes the conversation context that I can reuse in another conversation, across products.
I would love to get some feedback and suggestions.
Is it really so simple that all information in an LLM comes from the probability of each token based on the prompt? So for any prompt, there is a probability distribution to continuing (after) that prompt to generate text?
All structure of information comes from probabilities of tokens (so all structure and information processing is a side effect of token probabilities)? Or is there other stuff going on? I know reasoning models have extra stuff but let's put that aside for now.
I wrote clark-agent, a small Rust library for running LLM tool loops.
The loop is:
context -> model -> tool call -> tool result -> repeat
The parts I wanted typed were:
- transcript messages - tool calls - tool results - stream events - tool schemas
The model/provider boundary is a StreamFn trait. Tools implement AgentTool. There are hooks for things like context transforms, tool gates, observers, and follow-up messages.
cargo install gsd-meta-manager
TUI for: - GSD milestone and phase progress - git history - open milestones - backlog - browser for the GSD markdown artifacts - tmux support to quickly jump into the project's running claude instance
All feedback, improvements, PR and questions are welcome.