Launched this week
Branchmore
Track tokens and analyze sessions to find/fix agent problems
31 followers
Track tokens and analyze sessions to find/fix agent problems
31 followers
Branchmore improves productivity for BOTH agents and the people who use them. To start, it combines token tracking (like ccusage) and session recording into one observability tool. Then our analysis goes deeper, by proposing improvements in 3 areas: 1) token costs, 2) unsafe tool calls and 3) agent productivity. Our data visualizations let you explore and understand what your agents have been doing. And we generate code/markdown to convert problem signals into fresh context/instrumentation.
Hello, hello! This is my first Product Hunt, and I'd like to share Branchmore with you all.
I built Branchmore to make BOTH developers and their agents more productive. I know how much effort it takes to stay on top of coding agent best practices, because I was an engineering manager at the cybersecurity startup Permiso, where I worked on improving AI adoption for a lot of engineers.
The tooling changes fast, and most people can't keep up with both the best practices and the instrumentation. Agent metrics and logs take a lot of effort to dig through, and need to be framed well to be useful. And we needed automation to refresh our context and tooling, which quickly fell out of date.
On top of that, LLMs are prone to hallucinating, misunderstanding and taking shortcuts, so agents will keep making mistakes even as models improve. These are subtle problems that show up in token costs and sketchy behavior, and people need more accessible tools to understand and improve their harness.
Here are some of the things you can do with Branchmore:
Track tokens and your subscription usage to balance between cost and velocity.
Track your historical 5-hour / 7-day rate limit trends, so you have real data when it feels like a model is burning tokens twice as fast as the day before. This is a very unique feature of Branchmore.
Review the Unsafe Tool Calls page to see whether your agents are doing anything risky, so you can reduce the chance of leaked secrets or data loss.
Make improvements using the actionable Ideas on the Weekly Pulse page, or set Weekly Goals for habits that move key productivity metrics.
Run the CLI's "tune" command to automatically generate fresh codebase context (AGENTS.md, Agent Skills) based on the negative signals we've observed. We're currently working on delivering that as a PR.
Create Groups in the Community pages so you can see what models your friends are using, and how many tokens they're burning. We want to add things like badges so you can explore AI tooling together.
Some people call this a self-improving harness and build big code factories for large teams and enterprises. I wanted something lightweight that an individual engineer can set up quickly to shed light on their agents' inefficiencies and find ideas to fix them. Beyond tools to make agents self-improving, we also believe in the idea of people continuously learning. After all, people operate the agents and operation takes skill.
Setup is easy - it should only take 5-10 minutes. You create a web account, download the terminal tool, and initialize it so it installs the hooks into your coding agents. We currently support Claude, Codex and Cursor. Check out our documentation for more details, and join our Discord if you have questions.
We have a generous free tier, and have paid plans for power users that already spend more than $100 per month on AI or who want our advanced analytics.
I'd love to get feedback on the product and the pricing structure, and how I can help engineers build better with their coding agents! I appreciate your time.
Cheers, Andrew