Boost

Boost is a free command line tool from JFrog that cuts the tokens coding agents waste on noisy tool output. It sits between the agent and the terminal, compressing repetitive command output before it reaches the model while keeping failures, summaries and the clues the agent needs for its next step. It works with Cursor, Claude Code, Codex, GitHub Copilot and OpenCode on macOS, Windows and Linux, installs in one line, and connects to an agent with a single init command. It is built for developers and teams paying for heavy agent usage who want lower cost and more usable context without changing their workflow. Built in filters are aware of command flags, and teams can write custom TOML filters for internal CLIs and scripts or let the agent mine local Boost history to draft the most useful ones. MCP filters, in staged rollout, compress large JSON responses from MCP servers. BoostGraph, in early access, gives agents structural code search in place of long search and read loops. A local observability report breaks each session down by model, subagent, turn, tool call and token usage, and a GitHub integration can keep a live pull request comment with usage and estimated cost. JFrog describes Boost as free forever with no signup required. No paid plan is published on the site. JFrog publishes a benchmark in which cost per task fell from $0.319 at baseline to $0.276 with Boost, a 13.5 percent reduction, and measurement happens locally so session data stays on the developer's machine.