Boost is a lazy-loaded local CLI layer that truncates verbose terminal output and replaces it with semantic markers to reduce token consumption for coding agents, caching logs locally for on-demand inspection. The tool, incubated within JFrog, runs entirely locally and seeks community feedback on its heuristic.
When building long-running coding agents, terminal output is one of the fastest ways to poison a context window. If an agent runs an intensive build, an install command, or a massive search (grep/find), it easily generates hundreds of lines of raw log noise. The agent reads it once to check the status, but you continue to pay for those massive, redundant log tokens on every subsequent turn in the conversation loop. To fix this, we designed a lightweight CLI layer called Boost to act as a contextual gatekeeper. We wanted a way to lean out context without stripping the agent's ability to debug deep errors. The Heuristic: Lazy-Loaded Context Pointers Instead of feeding raw stdout/stderr back to the LLM agent, the CLI layer intercepts the stream: Noise Truncation: It identifies repetitive patterns, verbose progress bars (like npm or pip install states), and successful compilation blocks. Semantic Markers: It replaces the raw log text with a lightweight token marker: [Terminal Output Truncated: 450 lines of webpack build logs. Reference ID: log_9a3b. Status: Success] Lazy Hydration: The raw log is cached 100% locally. If the agent hits a roadblock later and explicitly decides it needs to inspect the compiler output to debug an error, it invokes a sub-command to "hydrate" that specific reference ID (e.g., boost inspect log_9a3b). In our internal testing, this drastically dropped token consumption per session without degrading task success rates. The Transparency Part (JFrog & Privacy) We operate as an internal startup incubated within JFrog. Being backed by a larger organization means we have the dedicated resources to build stable developer tooling, but we are taking this one step at a time. Let's be 100% transparent about where the tool stands right now: Privacy & Telemetry: It runs 100% locally on your machine and processes truncation entirely in-memory. Your raw terminal outputs, code, and logs never leave your machine. The only network call it makes is a lightweight telemetry ping to track aggregate token savings so we can measure the tool's real-world impact and improve it. The Code: The binary is completely free, but the source code is currently closed while we validate the core mechanics. You can check it out at boost.jfrog.com or find the releases on our GitHub page: github.com/jfrog/boost Beyond the tool itself, we want to know if this logic breaks down at scale. How are you keeping your agent context lean without breaking workflows? Are you using custom system prompts, vector embeddings for logs, or just letting the tokens burn?
Hugging Face redesigned its `hf` CLI to be optimized for both human users and AI coding agents like Claude Code and Codex, with agent-aware output rendering and benchmarking showing up to 6× token savings versus no-CLI baselines on complex tasks.
lean-ctx is an open-source Rust-based context runtime that reduces token costs for AI coding agents like Claude Code, Cursor, Copilot, and others by 60–95% through file read compression and shell output optimization. It operates as a Shell Hook and MCP Server with 56 tools and multiple read modes.
LangChain releases Deep Agents v0.7, a leaner agent harness that cuts base input tokens by 65% through removing the default system prompt, trimming tool descriptions, and making TodoListMiddleware opt-in, with comparable performance validated across multiple models.
LeanCTX is updated to act as a gateway managing tool outputs for AI agents, compressing large data, scrubbing secrets, and integrating external tools via MCP to prevent token blowup.