cd /news/ai-agents/show-hn-whole-body-compute-a-local-5… Β· home β€Ί topics β€Ί ai-agents β€Ί article
[ARTICLE Β· art-167998] src=github.com β†— pub= topic=ai-agents verified=true sentiment=↑ positive

Show HN: Whole Body Compute: A local 5-organ runtime for coding agents on Mac

Whole Body Compute released `body`, a local macOS runtime that intercepts coding-agent file reads and tool outputs in under 3ms before they reach cloud billing, claiming up to 98.7% token reduction on large reads and 96.9%–99.5% sensory token reduction on desktop vision. Across 146 live unscripted agent sessions, including 100 official princeton-nlp/SWE-bench_Verified runs across 10 Python organizations, the project reports 30/30 (100% Pass@1) and 10/10 Pass^3 with context falling from 203.08M to 137.23M tokens, which it describes as zero accuracy loss. The tool auto-wires PreToolUse/PostToolUse hooks for Claude Code, Cursor, Windsurf, Cline, Goose and Antigravity, and adds a 23-vendor secret redaction layer claiming 0 secret leaks and 0.0% ASR on AgentDojo.

read5 min views4 publishedNov 11, 2026
Show HN: Whole Body Compute: A local 5-organ runtime for coding agents on Mac
Image: Michielbdejong (auto-discovered)

Cut agent token burn and latency with zero accuracy loss on SWE-bench Verified.

Whole Body Compute makes AI computing simpler, faster, and local: giving Claude Code, Cursor, Windsurf, Cline, Goose, and Antigravity a <3ms AST context shield, native macOS vision, and on-device memory right on your Mac (with cross-platform Blood-Brain Barrier & FinOps audit support on Linux).

curl -fsSL https://wholebodycompute.com/begin | sh

uv tool install git+https://github.com/wholebodycompute/whole-body-compute.git && body begin

git clone https://github.com/wholebodycompute/whole-body-compute.git
cd whole-body-compute
python3 -m venv .venv && source .venv/bin/activate && pip install -e .
body begin

Or run a read-only 30-day token spend audit across your local agent logs with zero installation:

curl -sSL https://wholebodycompute.com/audit.py | python3

Cloud coding agents burn massive context windows re-ingesting full files, noisy test logs, and screenshots on every turn. Whole Body Compute (body) sits locally between your coding agent and macOS/Linux hardware, intercepting bloated reads and tool outputs in <3ms before they hit your cloud billing meter.

β”Œβ”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”
β”‚                            WHOLE BODY COMPUTE (5 CORE ORGANS)                          β”‚
β”œβ”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”¬β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”¬β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€
β”‚ 01. THE HEART            β”‚ 02. THE BRAINSTEM        β”‚ 03. THE EYES & HANDS             β”‚
β”‚ Thermal, Perfusion & CSI β”‚ Local Metal, Router & KV β”‚ Foveal AX DOM, Tactile & Vagus   β”‚
β”œβ”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”Όβ”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”΄β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€
β”‚ 04. THE HIPPOCAMPUS      β”‚ 05. THE IMMUNE SYSTEM (BLOOD-BRAIN BARRIER)                 β”‚
β”‚ Memory WAL, Flush & REM  β”‚ Pre/Post Hooks, SensoryVector, <3ms AST Slicer & SecGate    β”‚
β””β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”΄β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”˜
Capability Standard Unshielded Cloud Agent With Whole Body Compute ( body ) Measured Impact
Harness Wiring Manual per-editor rules & scattered scripts body begin auto-wiresPreToolUse /PostToolUse hooks across Claude Code, Cursor, Antigravity, Windsurf, Cline & Goose 60-second setup + 30-day FinOps receipt
File & Log Reads Unbounded read_file (2,000+ lines) & rawpytest /cat dumps Blood-Brain Barrier (<3ms AST Symbol Slicer) + local MetalReconScout Up to 98.7% token reduction on large reads
Desktop Vision Full-screen PNG loops ( 7,650+ tokens/step) or heavy 24/7 video recorders Foveal Vision (AXUIElement DOM) +<50ms Tactile git/file delta sensor (<50MB RAM) 96.9%–99.5% sensory token reduction
Session Hygiene Cognitive Sentry (body context) +StateCapsule Flush ( body flush)[Pro] Zero-loss handoff in<30 steps
Pre-Egress Security Cloud classifiers run after data leaves your machine <1ms 23-vendor secret redaction,1Password /Keychain blindfold,<2ms APFS CoW airlock & AST SecGate 0 secret leaks Β· 0.0% ASR onAgentDojo

Evaluated across 146 live unscripted agent sessions (including 100 official upstream princeton-nlp/SWE-bench_Verified runs across 10 Python organizations at untouched historical commits, graded strictly against held-out FAIL_TO_PASS + PASS_TO_PASS test patches):

Evaluation Suite Baseline (Unshielded) + Whole Body Compute Key Empirical Result
SWE-bench Verified (k=3, 60 Live Trials)Gemini 4 Carbon across10 official orgs 30/30 (100% Pass@1 )10/10 Pass^3 Β·203.08M ctx 30/30 (100% Pass@1)**** 10/10 Pass^3 Β· 137.23M ctx Zero accuracy loss (100% strict3/3 consistency),-32.4% total context (-65.85M tokens),-25.6% billed input ,-18.8% fewer turns
SWE-bench Verified (Claude Code 2x2, 40 Runs)Claude Sonnet &Claude Haiku CLI Vanilla Sonnet :9/10 ($4.00 )Vanilla Haiku :8/10 ($4.42 ) Sonnet + WBC: 10/10 ($3.89)**** Haiku + WBC: -36.1% ctx ($3.05) Sonnet + WBC solved 10/10 (100%) and cost**-12.0% LESS than Vanilla Haiku** ($3.89 vs.$4.42 ) while beating it by+20.0pp accuracy
Multi-Track Live A/B (18 Sessions)SWE-bench +CORE-Bench Hard +OSWorld Haiku 4.5 :9/9 (100% )5.26M ctx Β·$1.60 Β·312.9s Haiku 4.5 + WBC: 9/9 (100%)**** 1.55M ctx Β· $0.51 Β· 190.5s -67.8% API spend (3.11x efficiency) ,-71.9% billed input ,-39.1% faster wall-clock time
Security & Reliability (AgentDojo + CyberSecEval 3) 100.0% Attack Success (12/12 ) 0.0% Attack Success (12/12 blocked) 0.0% false-positive rate (6/6 benign passed),0.102ms in-process gate overhead

Full per-repository tables, 2x2 model matrices (Sonnet, Haiku, Gemini, GPT-OSS 120B), raw JSON telemetry receipts, and 1-command verification harnesses: benchmarks/README.md.

body begin          # Wire Blood-Brain Barrier hooks & print 30-day FinOps ROI receipt
body audit-spend    # Scan Claude Code, Cursor, Antigravity, Windsurf, Cline, Goose, Codex & Aider spend
body savings        # Inspect live Blood-Brain Barrier token savings & active shield status
body report         # 5-Pillar operational & FinOps scorecard (Session, 24h & 7d trends)
body doctor         # Sub-second 5-organ health check (--exercise runs closed-loop reflex verification)
body heart          # Thermal state, Unified RAM headroom & model perfusion status
body eyes           # Live macOS AXUIElement Foveal DOM snapshot + on-demand Apple Vision OCR
body hands          # Deterministic UI actuation (ax_act) & <50ms tactile git/file delta sensor
body immune         # Blood-Brain Barrier audit log (--window 24h --tools --blocked), APFS CoW & SecGate
body daemon status  # Manage portable macOS LaunchAgents (install / status / logs)

body license        # Activate Pro/Team Polar key (--activate <KEY>) or inspect tier quota
body context        # Context Saturation Index (CSI) & bloat attribution (--insights) [Pro]
body flush --auto   # Grounded StateCapsule session handoff & metabolic noise purge [Pro]
body memory         # Persistent Hippocampus SQLite WAL & FTS5 recall (--remember / --recall) [Pro]
body perfusion      # Evict idle local Metal models (--evict) or pre-warm weights (--prewarm) [Pro]
body sleep          # Nocturnal REM consolidation & Radix KV-Cache prefix digest [Pro]

Deep-dive component reference and glossary: docs/WIKI_AND_GLOSSARY.md.

Free forever for individual developers, researchers, and internal non-competing business use under the Open-Core Community License (LICENSE) and PATENTS.md - commercial resale, managed hosting, or embedding into competing IDEs, agent harnesses, gateways, or FinOps products requires a separate commercial license (hello@wholebodycompute.com).

── more in #ai-agents 4 stories Β· sorted by recency
── more on @whole body compute 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain β€” perfect for shipping the agent you just read about.

$git push zahid main
β†’ Live at https://your-agent.zahid.host βœ“
Get free account β†’ Pricing
from €0/mo Β· no card required
LIVE [news/show-hn-whole-body-c…] indexed:0 read:5min 2026-11-11 Β· β€”