How it works

#What Branchmore tracks

Tracking token costs and subscription rate limits is one of the most urgent needs Branchmore addresses, but it's important for engineers to reflect on the full range of problems they can run into when working with agents:

• AI cost (tokens and subscriptions). Burning tokens, tokenmaxxing, and "using multiple $200 subscriptions" are contentious topics, but they help inspire bold new ways to use AI. On the other hand, saving tokens is what makes the practice affordable.
• Productivity and velocity. Costs aren't everything. Are you using your time well? We look for disagreements and problems in your usage that are costing you time. We also consolidate your conversations with AI to help you visualize outcomes over time.
• Safety and risk. As we seek to make AI agents more productive, we often need to "automatically approve tool calls" that could be risky or destructive. We also run more agents in parallel, which compounds the risk. Branchmore inspects agent activity for risky actions such as unsafe installations, privilege escalation, and secret leakage.

#How to use Branchmore

You can think of Branchmore as covering "human learning" and "agent learning".

#Human learning

Human learning happens when you review our backend's analysis through the lens of cost optimization, engineering productivity, and cybersecurity. Here are a few ways you can use Branchmore:

• Track tokens and subscriptions. Your AI tools may show you basic rate limit and token usage, but you can learn so much more when you look at historical trends and break down granular, token-level information.
• Use the Weekly Pulse to find improvements. We synthesize your data into an opinionated set of Improvement items to help you find problems and fix them. Weekly Goals are thresholds you can set for yourself to build positive habits.
• Review Tool Calls to find unsafe commands. As engineers grant agents more permissions to improve their autonomy, we need better monitoring to tell whether those agents are taking risky actions.
• (Experimental) Ask questions about your agent's history. Call bmor inspect in your terminal. We package it as a local MCP server so you can combine Branchmore's analysis with your local codebase without granting us access to your code.

#Agent learning

Agent learning involves generating code and context for your agents. Our backend analyzes token burns, hallucinations, and disagreements to find where your repositories have fallen out of date. We automate the documentation and instrumentation so your agents self-improve over time. This is our version of a "self-improving harness" or a "software factory".

• (Experimental) Improve your AGENTS.md. Call bmor tune [topic] in the CLI. This launches your agent of choice not only to enforce AGENTS.md best practices, but also to draw on the many signals we gather to suggest more specific improvements.

#Related