Agy Fast
Non-invasive MCP accelerator plugin for Google Antigravity (CLI/Desktop/IDE): cuts LLM turns (-34%), token quota (-31%), and upload traffic (-40%) without affec…
Latency and Tool-Call Accelerator Plugin for Google Antigravity
Zero runtime dependencies · Single-file ESM bundles · Policy-compliant · Works on Antigravity CLI (agy), Desktop 2.0, and IDE
English · 简体中文 · Changelog · Latency Analysis Report · 性能分析报告(中文脱敏版)
Note
Non-invasive — zero impact on Antigravity's native UI or behavior: agy-fast runs strictly as a standard local MCP plugin and never alters, disables, or interferes with Antigravity's built-in features — including the diff viewer/window, modified-files list, checkpoints, native tools, or safety checks.
Why agy-fast Exists
Analysis of 55 real Antigravity sessions (8,446 LLM turns, 17,302 tool calls, 55.25 hours of single-agent active time) showed that slow tasks are rarely caused by slow token generation:
- Per-turn Time to First Token (TTFT) floor (~6.1 s median): 68.1% of turns emit fewer than 700 tokens, so roughly 79% of a median turn is spent waiting for the first byte. Every unnecessary round-trip turn costs ~6–8 seconds.
- Serial tool execution and fragmented edits: 52.5% of
edit_filecalls are extra same-file splits in the same turn, and 46.3% ofrun_commandcalls are read-only searches (rg/grep/find/git status) that pay shell and remote-check overhead (~2.0 s median vs. 2.1 ms for localrg). - Cold-start exploration: The first 3 turns of a session average 49.2 seconds inspecting directory layout, git state, and project scripts.
- Context growth and cache misses:
read_fileaccounts for 81.9% of all tool-returned bytes; gaps over 60 seconds raise prompt cache miss rates from 1.4% to 15.9%–93.8%.
Read the full sanitized data breakdown in docs/ANALYSIS_REPORT.md.
agy-fast addresses these bottlenecks strictly within Antigravity's official plugin specification: it reduces round-trip turns, merges repeated tool calls inside a persistent local Stdio MCP server, keeps context compact, and injects a tiny startup snapshot once per session.
Real-World Impact (Sanitized Telemetry)
Across 10 days of real coding transcripts (14 baseline main sessions / 5,771 turns vs. 7 agy-fast main sessions / 901 turns, plus 104 baseline vs. 4 agy-fast subagent sessions; all repository names, paths, and session identifiers stripped), each agy-fast call batched an average of 4.89 logical operations (5.95× in subagents) and delivered measurable savings across latency, quota/token cost, and upload bandwidth:
| Metric (normalized per logical action) | Main Sessions (Baseline) | Main Sessions (agy-fast) |
Delta | >=150k Long Context |
Subagents |
|---|---|---|---|---|---|
| LLM turns / 10 actions | 5.23 | 3.47 | -33.7% | -38.2% | -58.8% |
| Upstream requests / 10 actions (LLM + remote checks) | 9.82 | 6.05 | -38.4% | -42.6% | -62.0% |
| Prompt Cache hit rate (miss spikes) | 89.6% (7.5%) | 94.0% (4.8%) | +4.4 pp | 95.0% (3.3%) | 92.4% (2.8%) |
| Uncached fresh input / action | 10,413 tok | 4,822 tok | -53.7% | -61.5% | -51.3% |
| Est. billable quota / action (1 : 0.1 : 5) | 22,458 eq | 15,387 eq | -31.5% | -36.5% | -54.0% |
| Tool-return upload / action (actual local bytes) | 5.06 KB | 3.05 KB | -39.7% | -42.6% | -13.3% |
| Incremental text upload / action (fresh input est.) | 35.6 KB | 16.5 KB | -53.7% | -61.6% | -51.3% |
| Full-context text upload / action (stateless resend est.) | 342.2 KB | 276.2 KB | -19.3% | -21.9% | -58.9% |
From the project's README.