Introducing Boost

Today, JFrog is releasing Boost: a free CLI that sits between your coding agents and the shell, compressing noisy command output before it reaches the model.

Install in one line and wire it into Cursor, Claude Code, Codex, Gemini CLI, or OpenCode — your agent's shell commands start saving tokens automatically.

What Boost does

Coding agents spend a large share of their context window reading terminal output — build logs, test results, git diff, Docker layers, and the rest. Much of that output is repetitive noise the model does not need to reason over.

Boost addresses that at the source:

• Save tokens — declarative TOML filters compress tool output before it enters agent context. In typical sessions, Boost cuts 60–90% of log tokens on common commands like git, test runners, and linters.
• Stay transparent — filtered output looks like a shorter version of the real command's output, not a new format. Flag-aware filters respect explicit verbosity (--nocapture, -la, etc.) so agents still get detail when they ask for it.
• Recover when needed — compression is reversible. When Boost hides something the agent actually needs, boost retrieve <id> restores the original output locally, with a short reason captured for learning.
• Never block — if a filter fails, Boost falls back to raw output. Hooks always exit 0 so the agent's command still runs.

Under the hood, Boost's rewrite hook turns supported commands into (cmd) | boost: the real tool runs unchanged; only stdout is piped through Boost on the way back to the agent.

Benchmarks: Terminal-Bench 2.0

To measure whether Boost helps agents without hurting their ability to complete real work, we ran Terminal-Bench 2.0 — a benchmark of hard, human-verified terminal tasks in containerized environments — with and without Boost enabled.

Setup

Benchmark	Terminal-Bench 2.0 (89 tasks)
Evaluated	81 tasks — 8 tasks excluded due to infrastructure constraints in our Harbor evaluation environment
Agent	Claude Code
Model	Claude Haiku 4.5 (us.anthropic.claude-haiku-4-5-20251001-v1:0) via Amazon Bedrock
Baseline	Agent runs commands with no Boost hooks
Boost	Same agent and model, with Boost hooks installed ((cmd) | boost rewrite + filters)

Harbor Hub results

Baseline Vs Boost

These links are public, but Harbor Hub requires you to sign in before you can view a job.

Task success (primary metric)

Boost preserved task completion on this benchmark: identical pass rate to the baseline.

Cost and token efficiency

Aggregate across all 81 tasks. Bars are relative to the baseline (=100%).

On long-horizon agent benchmarks, most tokens are reasoning and context — not raw tool stdout — so aggregate input-token savings are modest even when per-command compression is much larger. The ~12% cost reduction on the same pass rate is the more meaningful signal: Boost makes the same agent work cheaper without changing how many tasks it completes.

For day-to-day development, per-command savings are much larger. In a typical 30-minute session (~60 shell commands), Boost reduces context from tool output by roughly 91% on common workflows (git, test runners, linters, Docker).

Safety and agent guardrails

Boost is an output filter, not a command interceptor that hides what the agent is doing.

• We moved from wrapping the command like boost pytest … | grep … to (pytest … | grep …) | boost — the agent still sees the full pipeline it proposed, and Boost only shapes the final stdout. That made it possible to measure accurately the bytes we save from the context window without interfering when the agent pipes or filters output itself.
• Boost does not bypass an agent's security model, permission prompts, or sandbox guardrails. It runs after the agent has already chosen and been authorized to run the command; it only shapes stdout on the return path.
• If filtering would hide important detail, Boost can append a boost retrieve <id> marker so the agent can recover the original locally — a deliberate, visible rollback rather than silent data loss.
• Hooks are designed to never block: malformed input, unknown commands, and filter errors fall back to unmodified behavior.

Boost compresses what the agent reads. It does not change what the agent is allowed to run.

How it fits your stack

Coding agents — boost init installs rewrite hooks for Cursor, Claude Code, Codex, Gemini CLI, OpenCode, and more.

Observability — boost report -w opens a local dashboard of commands, token savings, and retrieve events.

Extensibility — you (or your agent) can easily identify noisy output and write custom TOML filters that save tokens for your special use cases.

Evaluations summary

Terminal-Bench runs used Claude Code with Claude Haiku 4.5 on Amazon Bedrock. Eight of 89 Terminal-Bench 2.0 tasks were excluded from this evaluation due to infrastructure limitations in our Harbor environment. Per-command token savings (60–90%) come from measured filter performance on real tool output; agent-benchmark aggregates include model reasoning tokens and understate per-command compression.

Availability and pricing

Boost is free — no signup, no credit card, no trial timer. The token-saving CLI is intended to stay free.

Install from GitHub or JFrog's distribution:

Docs and a walkthrough demo live at boost.jfrog.com. Source is on GitHub.