JFrog: Boost Enters Public Preview

What it claims

Boost is now in free public preview at boost.jfrog.com. The specific waste figures the summary omits: input context is ~70% of the total AI bill, and per JFrog's tracing over 90% of ingested tokens are read once and discarded. Individual Cursor bills hit >$700 for under 2,000 requests; single Claude Code sessions exceeded $1,700 — the pain that triggered the project.

How it works

Boost wraps agent shell commands with a boost prefix (boost ls -l, boost kubectl apply -f ...) and intercepts output before it reaches the model. Three named waste patterns, with the page's numbers:

• Test log avalanche: an agent was fed 847 lines of a pytest log when only the 9-line failure traceback mattered; the log-distiller shrinks such payloads by >80%.
• Blind grep/find: a getUser query returned 119 unordered matches (minified caches, docs, mocks) that the agent re-read every turn; Boost replaces this with an LSP-backed binary giving language-aware search.
• Truncation hallucination loop: naive slicing caused agent amnesia — re-running the same ls/cat up to six times to re-verify state, burning more than it saved.

The shipped product is two capabilities rather than a blind filter:

• boost skill — a minimal system prompt that tells the agent what Boost is and how to interact with the optimization layer, so it isn't silently manipulated.
• boost retrieve — fetches the full unedited output of the last command from session history when the agent needs filtered data back.

Evidence

Rolled out gradually across JFrog's 1,000+ engineer R&D org (Cursor and Claude Code), scaling from a few billion to over 100 billion tokens saved while agents stayed faster and more accurate. JFrog states it tracked boost retrieve frequency and per-technology context needs to guide trimming. These are self-reported internal figures with no external baseline.

Limits

• Self-reported vendor numbers from one organization; no independent benchmark.
• The boost-prefix design created a real security flaw JFrog flagged: one "Always Allow" click on boost ls -l grants blanket approval to everything including boost rm -rf / — they say permission handling was reworked, but verify how before granting approvals.
• Public preview, not GA.
• Telemetry claim to check: JFrog states it collects only high-level data (token counts, macro errors) and never sees, stores, or transmits code or conversations.