JFrog releases Boost CLI, claims 13.5% coding-agent cost reduction

• JFrog's free Boost CLI compresses shell output and MCP responses before they enter a coding agent's context, according to hasantoxr's hook description.
• The launch names Claude Code, Cursor, Codex, and GitHub Copilot as supported surfaces in hasantoxr's supported-agent list, while JFrog's current overview also lists Gemini CLI and OpenCode.
• JFrog reports Terminal-Bench 2.0 cost per task fell from $0.319 to $0.276, a -13.5% change, in hasantoxr's Terminal-Bench post.
• The “same success rate” wording needs a narrower reading: hasantoxr's Terminal-Bench post describes 89 tasks, while JFrog's Performance Lab report puts its matching 30.9% pass rate in a separate 81-task check.

The official homepage's test-runner demo keeps the failing assertion while dropping the routine passes. Its Performance Lab documentation says explicit verbose flags preserve detail, hooks fail open, and a local retrieval command can restore the original output.

Boost installs with a shell command, then boost init connects an agent. JFrog describes the tool as free with no signup in its quickstart.

• curl -fsSL https://boost.jfrog.com/install.sh | bash
• boost init
• Supported by the current docs: Cursor, Claude Code, Codex CLI, Gemini CLI, OpenCode, and GitHub Copilot.

Boost hooks append themselves to the command path as (cmd) | boost, per JFrog's technical documentation. The agent's command and any existing pipes run first, then Boost filters the final stdout on its return to context.

• Permissions: Native prompts still see the executable and pipeline the agent requested.
• Failure handling: Unknown commands, filter misses, and hook errors fall back to raw output.
• MCP: PostToolUse hooks can re-encode JSON responses as TOON before they reach context.

Boost has three compression surfaces.

1. CLI Filters: JFrog says noisy tools such as docker build, npm install, test runners, and linters typically shrink 60% to 90% after routine log lines are removed, while failures and summaries remain.
2. Boost Graph: The homepage example takes a grep-plus-read loop from 2,400 tokens to 1,000, a -58.3% reduction, by returning a definition and path rather than a file wall.
3. MCP Filter: A TOON example re-encodes roughly 65 JSON tokens as 29 TOON tokens, a -55% reduction.

JFrog's benchmark chart reports cost per task across the 89-task Terminal-Bench 2.0 suite.

• Baseline → RTK: $0.319 → $0.339, +6.3%.
• Baseline → Boost: $0.319 → $0.276, -13.5%.
• RTK → Boost: $0.339 → $0.276, -18.6%.
• Full-suite duration: baseline → Boost, -28.9%, according to JFrog's Performance Lab report.

The pass-rate result comes from a separate 81-task check: 30.9% → 30.9%, 0.0 points, in the same report. That scope is more specific than the 89-task, same-success-rate wording in hasantoxr's benchmark post.

boost report lists the lines compressed and the reason for each, and filtering can be disabled for an individual command.

JFrog's token-savings documentation says the report tracks raw versus filtered size, estimates tokens saved per command, and aggregates session savings. boost retrieve <id> restores the locally stored original output; JFrog says filtering happens locally and native permission prompts stay unchanged.

Compression can hide detail the agent later needs. hasantoxr's control post describes a per-command off switch, while JFrog's Performance Lab documentation says repeated retrievals can automatically disable a bad filter.

The early integration record also contains two resolved bugs. An OpenCode issue found that boost init --opencode installed a skill file where OpenCode did not discover it, leaving Boost-native commands unavailable to the agent; a maintainer closed it with a v0.9.15 fix. A separate GitHub Actions bug tied intermittent Go go vet failures to Boost forcing GOPROXY=direct; it was also closed with a v0.9.15 update.