GitHub - diegosouzapw/OmniRoute: Never stop coding. Free MIT AI gateway: one endpoint, 352 providers (150+ free), 1200+ models Kimi, Claude,

🚀 OmniRoute — The Free AI Gateway

💰 ~1.51B Free Tokens / Month

Stacking free tiers by hand is painful — dozens of SDKs, dozens of rate limits, and no idea how much you actually have.

OmniRoute catalogs 446 free-tier entries across 38 recurring pool keys and computes the token headline from the 20 pools with a published positive monthly budget , deduplicated by shared pool.

The result stays visible on the dashboard ( /dashboard/free-tiers ).

Animated summary of the live /dashboard/free-tiers page. Full methodology (pool dedupe, credit tiers, provider terms): docs/reference/FREE_TIERS.md .

These figures are re-audited every two weeks against the live catalog and move both ways — a provider ends a free tier and the number drops; a new one lands and it climbs. We publish what the catalog actually computes, never a rounded-up best case.

⭐ Star the repo if OMNIROUTE helped you save money and make your work easier.

💬 Join the community

👋 Follow the maintainer — get new providers, releases & tips first:

Questions, provider tips, roadmap & support → Discord · Telegram · WhatsApp 🌍 Global / 🇧🇷 Brasil / Portal

📈 The Gateway Keeps Growing

v3.8.49 v3.8.50 v3.8.51+

🌐 Providers 290 352 more queued

🧠 Unique chat model IDs 1185 1312 —

🖼️ Modality Bridge — 🆕 vision + audio + video —

📡 Radar free catalog — 🆕 opt-in —

⚖️ Quota-aware scheduling — 🆕 Quota-Share —

📊 Quota telemetry — 🆕 live —

→ Roadmap — riding the rail to v3.9.0 LTS

🧩 Available

🚀 Start 🚀 Quick Start 📦 Install 🆓 Zero-config

💡 Learn 💥 The Promise 🤔 Why OmniRoute 🏆 What Sets Apart

⚙️ Features 🎯 Combos 🌐 Providers 🔌 CLI & MCP

🗜️ Compression 🖥️ Where It Runs 🔒 Private

👀 See it 🎬 In Action ✨ What's New 🤖 Compatible CLIs

💚 Support 💚 Support / Donate 💬 Community 💖 Sponsors

📦 Project 🛠️ Tech Stack 📖 Docs 👥 Contributors

🌐 In 43 languages

🆓 Works the second you install it — no keys, no config

Fresh install, zero credentials — auto already works:

curl http://localhost:20128/v1/chat/completions \

-H " Content-Type: application/json " \ -d ' {"model":"auto","messages":[{"role":"user","content":"Hello!"}]} '

Prefer a specific free backend? Call oc/… (OpenCode Free) directly. Then graduate to auto and let OmniRoute pick.

📦 Copy-paste quickstart scripts for Python, Node.js, PHP, and cURL → examples/quickstart/

💥 The Promise

🤔 Why OmniRoute?

🤝 Supported by our Open Source Friends

Want to join as an Open Source Friend? These are the companies that back open source and help keep OmniRoute moving — and we say publicly where every token they give us goes. Reach out: [email protected]

Kimi Moonshot AI

Thanks to Kimi (Moonshot AI) , our founding Open Source Friend, for backing this project!

Kimi is the AI lab behind the open-weight K2 and K3 model families — Kimi K3 delivers a 1M-token context window, native vision and frontier-level coding at a fraction of closed-model prices, and works out of the box with Claude Code, Codex and every coding tool OmniRoute serves.

What Kimi's support powers: Kimi's API credits power OmniRoute's AI-validated release pipeline — the merge validation powered by Kimi K3 stage that reviews every pull request before it ships — plus day-to-day feature development.

First-class Kimi support ships on both rails: the direct Kimi API ( kimi-k3 ) and the Kimi Code coding plan (OAuth and API key). OmniRoute is also the first Brazilian open-source project in Kimi's support program. Get a Kimi API key with 15% extra credits →

Cheaper Inference cheaperinference.com

Thanks to Cheaper Inference , an OmniRoute Open Source Friend, for backing this project! Cheaper Inference is a cost-ranked gateway that resells 42 frontier models — Claude, GPT-5.

x, Gemini, Kimi K3, GLM, DeepSeek, Grok and MiniMax — behind one OpenAI-compatible endpoint, routing each request to the cheapest eligible provider without ever charging above the model maker's list price.

First-class support in OmniRoute: Chat Completions, the native /v1/responses endpoint, vision, tool calling and 3 image models ( grok-imagine , nano-banana-pro , nano-banana-2 , reachable as cheaperinference/<model> ). Get an API key →

Links tagged aff=omniroute are partner links. They fund the project at no extra cost to you.

🎟️ Affiliates Promo — free signup coupons from providers we don't sponsor (click to expand) This section is for referral/coupon codes only. Sponsored partnerships live in 🤝 Supported by our Open Source Friends above.

OmniRoute has no sponsorship or partnership with the providers listed here — these are public coupons anyone can use.

AgentRouter agentrouter.org

AgentRouter — affiliate signup · $100 free credits on signup (free server, expect higher latency — best for testing, not production). First-class support in OmniRoute since v3.8.

50 : Chat Completions, the Anthropic-compatible wire format and the OpenAI-compatible path. Available models include claude-opus-4-8 , claude-opus-5 , gpt-5.6-sol and more. Grab your $100 →

⚠️ Affiliate link — OmniRoute has no sponsorship or partnership with this provider.

Know another provider with a generous free signup coupon that benefits OmniRoute users? Open an issue and we'll add it here.

🎯 Combos — The Flagship

A combo is a chain of models OmniRoute routes across automatically . If quota runs out, a provider fails, or costs spike, the combo can move to the next eligible healthy model. 🛡️

⚡ Zero-config — just use auto

No combo to create. Set your model to auto (or a variant) and OmniRoute builds a virtual combo from your connected providers, scored live:

Model ID What it optimizes for

auto 🎯 Balanced default (LKGP — sticks to your last good provider)

auto/coding 🧑‍💻 Quality-first weights for code generation

auto/fast ⚡ Lowest latency first

auto/cheap 💰 Cheapest per token first

auto/offline 🔋 Most quota / rate-limit headroom first

auto/smart 🔭 Quality-first + 10% exploration to discover better models

auto/lkgp 📌 Explicit last-known-good-provider stickiness

auto/chaos 🧪 Fault-injection weights for resilience testing (chaos engineering)

🔀 Or build your own — 19 routing strategies

All 19 strategies — mix & match per combo step:

# Strategy What it does

1 priority First-target ordered list — drain each before the next 🥇

2 fill-first Fill each target's quota fully before moving on

3 weighted Weighted random by per-target weight

4 round-robin Cycle through targets in order

5 p2c Power-of-two-choices random load balancing

6 least-used Pick the target with the lowest current load

7 random Uniform random pick (deduplicated)

8 strict-random Random without de-duplicating repeats 🎲

9 cost-optimized Minimize $ per request from live catalog pricing 💸

10 headroom Pick the target with the most remaining quota

11 reset-window Prefer the target whose quota window resets soonest

12 reset-aware Rank by quota reset time — short windows first 📊

13 context-relay Hand off context across targets for long conversations 🧠

14 context-optimized Pick the best fit for the current context size

15 cache-optimized Pin each reusable prompt prefix to the same account — maximize prompt-cache hits 🎯

16 lkgp Last-Known-Good Path — pins to the last successful provider, then falls back to rules

17 auto 16-factor live scoring across every connection 🤖

18 fusion Fan out to a panel of models + a judge synthesizes one answer 🧬

19 pipeline Chain steps — each target's output feeds the next one 🔗

The Auto-Combo engine scores every candidate on 16 factors (health, quota, cost, latency, task fit, quality, session availability…) — see docs/routing/AUTO-COMBO.md .

🧱 Resilience is built in (3 independent layers)

📖 Auto-Combo Engine · Resilience Guide

🏆 What Sets OmniRoute Apart

📊 Full methodology & per-feature detail vs 9router, OpenRouter, CLIProxyAPI & LiteLLM → docs/comparison/OMNIROUTE_VS_ALTERNATIVES.md

💚 Support OmniRoute

OmniRoute is MIT-licensed and maintained in the open. If it saves you time or money, here's how to keep it independent — pick whatever fits you. Sponsorship never affects routing priority; it buys visibility, not ranking.

⭐ Star the repo Free — genuinely helps visibility Star OmniRoute

🐙 GitHub Sponsors One-off or monthly · zero platform fee github.com/sponsors/diegosouzapw

☕ Ko-fi Quick one-off tip, no signup for the donor ko-fi.com/diegosouzapw

🧋 Buy Me a Coffee Small, informal gesture buymeacoffee.com/diegosouzapw

🖐 Liberapay Recurring · non-profit · open source liberapay.com/diegosouzapw

🇧🇷 PIX (Brazil) Instant, no fees key & QR below

₿ Crypto BTC · ETH · USDT-TRC20 · USDC-Solana addresses below

🇧🇷 PIX — instant, no fees (Brazil)

Key (random): 5d865059-bc44-483a-962d-43ceb80126eb

Pix copia-e-cola:

00020101021126580014br.gov.bcb.pix01365d865059-bc44-483a-962d-43ceb80126eb5204000053039865802BR5922OMNIROUTE CONTRIBUICAO6006BRASIL62070503***630475DD

₿ Crypto — BTC · ETH · USDT-TRC20 · USDC-Solana (click to expand)

₿ BTC Bitcoin (SegWit) bc1qh00smz004sy85wyl28v77tenkt3ckl6eaep7fd

Ξ ETH Ethereum (ERC20) 0x64Cf6B68A6Ff34288e89172950a2d00102337a84

₮ USDT Tron (TRC20) TKAF41JpuQrHbKTnsQa9svJE2T192Hvsc2

$ USDC Solana 2emNNZzVVWQc3FQ2wk9M6qXUQmW8AKdjjL174fXR28Tu

⚠️ Send each coin only on the network shown — sending on the wrong network can lose the funds.

🐛 Found a bug or have feedback? Open a Discussion .

Developer notes: The project may generate a local .env file during npm install/postinstall for developer convenience. This file is intentionally ignored via .gitignore (see .

gitignore ) and must never be committed — if accidentally committed, rotate any exposed secrets and remove the file from history. See docs/DEVELOPER-ENVIRONMENT.md for guidance on managing local environment files and secrets.

📡 OmniRoute Radar

The main free-tier headline remains ~1.51B tokens/month from the documented, pool-deduplicated catalog above. Temporary provider signup credits can separately lift the first month to ~2.13B .

Radar is an optional, signed catalog overlay for people who want fresher free-model availability between OmniRoute releases; the community catalog and every existing free feature remain free.

Supporters can receive the live catalog and additional provider opportunities. Its separate, mutable ceiling is approximately 3B tokens/month at most , depending on provider availability.

That ceiling is not a guarantee: providers can change quotas, eligibility, models, or regions at any time.

Radar is opt-in and GET-only. The OmniRoute client does not upload prompts, traffic, provider configuration, usage telemetry, or local announcement-dismiss state. Learn about eligibility and the current catalog at radar.omniroute.online/planos .

✨ What's New

Recent highlights from v3.8.20 → v3.8.50 . Full history in CHANGELOG.md .

• 🎛️ OmniConductor — inbound A2A delegation to your agent fleet, Conductor skills on the Agent Card, and a dashboard panel with Faro push-to-talk voice chat. → A2A Server

• 🛂 Adaptive admission & overload protection — heavyweight chat requests queue instead of 503ing, with atomic RPM rolling leases per connection. → Resilience Guide

• 🗂️ Canonical /v1/models ordering — one contiguous provider-grouped block per provider (combos pinned first), stable across every catalog source. → API Reference

• 🗜️ Compression hardening — default-on inflation guard, Caveman packs for DE / FR / JA + Chinese (wényán), RTK filters for Gradle & .NET. → Compression

• 💸 Honest flat-rate cost — subscription / coding-plan providers read $0 in cost analytics; budget, quota & routing keep estimating. → API Reference

• ⚖️ Quota-Share routing — split a shared account's quota fairly across pooled keys, work-conserving so idle slices are lent out. → Resilience Guide

• 🤖 One-command CLI/agent setup — 13 registered setup-* commands; omniroute run launches 7 CLIs (Claude Code, Codex, Aider, Goose, OpenCode, Qwen Code, Gemini CLI); omniroute configure supports 10 targets with an interactive provider+model picker and per-context favorites. → CLI Integrations

• 🛰️ Remote mode — drive a remote OmniRoute with scoped tokens ( connect / contexts / tokens ) + an antigravity OAuth helper for VPS installs. → Remote Mode

• 🧭 Smarter auto-routing — auto/<category>:<tier> combos, Fusion (model panel + judge), task-aware routing, per-request model / mode / USD-budget overrides. → Auto-Combo

• 🗜️ Pluggable compression — 12 composable engines + Compression Studios: LLMLingua-2, two-tier Ultra, omniglyph, per-step fidelity gate, GCF v3.2, drag-reorder editor. → Compression

• 🕵️ Transparent MITM decrypt (TPROXY) — capture CLIs that ignore proxy env vars, with a per-SNI CA + trust-store installer. → MITM/TPROXY

• 💸 Cost telemetry everywhere — X-OmniRoute-* cost/usage headers on every endpoint, cache-HIT savings header, per-key USD spend quotas. → API Reference

• 🧠 Memory you control — off by default, opt-in int8 vector quantization + typed decay, per-request x-omniroute-no-memory . → Memory

• 🛡️ Security — prompt-injection guard on every LLM route (red-team suite), opt-in credential-masking guardrail (redacts leaked API keys/secrets in both directions), free DuckDuckGo last-resort web search, and an optional OIDC login gate for the dashboard (password login always stays available). → Guardrails

• 🖼️ New endpoints — /v1/ocr (Mistral OCR) and /v1/audio/translations (Whisper-style) round out the media surface. → API Reference

• 🎨 Image / video / audio generation — one API for media: xAI Grok Imagine & Novita AI video, ComfyUI, Magnific, Adobe Firefly, Segmind, and speech providers such as ElevenLabs. → API Reference

• 🌍 Deployment & ops — reverse-proxy basePath , browser-language auto-detect, per-key device tracking, root-less MITM trust, zh-TW localization. → Environment

• 🤝 More providers & agents — cloud agents (Codex Cloud, Cursor, Devin, Jules), Grok Build (xAI) with browser + OAuth login, Ollama first-class card, Claude Opus 5 & Sonnet 5, Kimi official partnership (Code/Web/Moonshot), Zed, Requesty, SenseNova, Yuanbao, Agnes AI… and a refreshed 352-provider catalog . → Providers

• 📡 Routing transparency — every response carries an X-OmniRoute-Decision header naming the strategy/provider/latency that served it, a new cache-optimized combo strategy + Auto-Combo cacheAffinity factor route repeat requests back to the connection holding the cached prefix, and a read-only /v1/auto-combo/{channel}/candidates endpoint exposes an auto/* channel's live candidate pool. → Auto-Combo

• ⚡ Local performance & infra — one-click local Redis, Cloudflare Workers / Deno Deploy relay deployers, Bifrost & Mux as supervised embedded services. → Embedded Services

• 🧩 Also in the box — plugin framework + marketplace, Omni/Agent/GitHub skills frameworks, Obsidian vault integration (22 MCP tools), OpenAI-compatible Batch & Files APIs, semantic response cache, gamification with leaderboards, ACP agent discovery (15 built-in agents), scheduled log export to BigQuery, auto/chaos fault injection, a Telegram bot bridge, an in-app version manager and LMArena-ELO free-provider rankings. → Docs

🤖 Compatible CLIs & Coding Agents

One config — http://localhost:20128/v1 — and every AI IDE or CLI runs on free & low-cost models.

Claude Code

Codex CLI

Cline

Kilo Code

Zoo Code

Continue

Aider

ForgeCode

jcode

DeepSeek TUI

CodeWhale

OpenCode

Factory Droid

Copilot CLI

Cursor CLI

Smelt

Pi

Grok Build

Hermes Agent

OpenClaw

Goose

Open Interpreter

Warp AI

Agent Deck

＋ also works with · Kiro · Command Code · Antigravity · Windsurf · AMP · any OpenAI-compatible tool

📖 Per-tool setup for all 36 tools (26 CLI Code's + 10 CLI Agents) → docs/reference/CLI-TOOLS.md · 🧩 OpenCode plugin → @omniroute/opencode-provider

Launch any supported CLI through OmniRoute in one command — no config files written, credentials injected per process, Qwen/Gemini get a throwaway isolated home:

omniroute run claude --model openai/gpt-5.4 # Claude Code omniroute run codex --model glm/glm-5.2 # OpenAI Codex CLI omniroute run aider --model glm/glm-5.2 -- --message " reply OK " omniroute run goose --model glm/glm-5.

2 omniroute run opencode --model glm/glm-5.2 -- run " reply OK " omniroute run qwen --model glm/glm-5.2 -- -p " reply OK " omniroute run gemini --model glm/glm-5.2 -- --skip-trust -p " reply OK "

Or pick provider+model interactively and write the tool's own config:

omniroute configure codex # also: claude opencode qwen aider goose gemini cline continue kilo

Every command honors the active remote context ( omniroute connect <host> ), --dry-run previews the exact env/args without executing, and --api-key-env NAME keeps secrets out of your shell history. → CLI Integrations

🌐 352 AI Providers — 152 Catalog-Marked Free

352 registered providers across the canonical chat, media, search, local, cloud-agent and system collections, including 152 carrying hasFree: true discovery metadata .

The chat model registry covers 229 providers / 2,554 distinct provider-model pairs / 1,283 raw model IDs ; the separate free-budget catalog has 446 per-model rows , 38 recurring pools and 53 recurring/keyless free-forever providers .

These are different denominators by design; definitions and pool-deduped calculations live in the Provider Reference and Free Tiers .

🏢 Every major lab — through one endpoint

OpenAI

Anthropic

Gemini

xAI Grok

DeepSeek

Mistral

Qwen

Meta Llama

Groq

NVIDIA

MiniMax

Cohere

Perplexity

HuggingFace

Together

Fireworks

Cloudflare

Baidu

…and 330+ more — every icon resolves live from the dashboard's provider catalog. 📖 Provider Reference

🆓 Free Forever — $0, no card

OpenCode Zen DeepSeek V4, Nemotron 3 No token cap

Kilo Code Auto-router, Tencent Hy3 Free forever

Requesty GPT-OSS 120B, Nemotron Free forever

SiliconFlow DeepSeek V3.2 / R1 Free tier

Z.AI GLM GLM-4.7 / 4.5-Flash Free forever

Baidu ERNIE ERNIE 4.0 Free forever

Qoder AI Qwen3-Max, Kimi-K2 Unlimited FREE

Pollinations GPT, Llama, Claude No key needed

Cloudflare AI 50+ models 10K neurons/day

NVIDIA NIM GLM, MiniMax ~40 RPM free

Cerebras GLM 4.7, GPT-OSS 1M tokens/day

OpenRouter :free models +$10 → higher RPM

📖 Full machine-readable catalog → docs/reference/PROVIDER_REFERENCE.md

🖥️ Where OmniRoute Runs — Anywhere

Same app, your machine, your rules. From a global npm install to your phone via Termux.

Platform Install Highlights

📦 npm (global) npm install -g omniroute One command, any OS

🐳 Docker docker run … diegosouzapw/omniroute Multi-arch AMD64 + ARM64

🖥️ Desktop (Electron) npm run electron:build Native window + system tray — Windows / macOS / Linux

💪 ARM native arm64 Raspberry Pi, ARM servers, Apple Silicon

📱 Android (Termux) pkg install nodejs && npx -y omniroute Runs on your phone , 24/7, no root

📲 PWA "Add to Home Screen" Fullscreen, offline, installable from browser

🧩 OpenCode plugin @omniroute/opencode-provider Native OpenCode integration

🤖 VS Code Copilot Chat install OmniCopilot extension Every OmniRoute model in the native Copilot Chat picker — stable & Insiders

🛠️ From source npm install && npm run dev Hack on it, contribute

📖 Docker Guide · Desktop · Termux · PWA · OpenCode

🧩 New: OmniRoute inside VS Code's native Copilot Chat

No new sidebar, no new chat UI — every model OmniRoute serves shows up right in the Copilot Chat model picker you already use . Since VS Code 1.122, provider models work without a GitHub sign-in or a Copilot subscription — agent mode, tool calling and vision, for free.

Install the OmniCopilot extension, point it at your OmniRoute server (defaults to localhost:20128 ), then open Copilot Chat → model picker → Manage Models… → OmniRoute .

Store Link Works with

🧩 VS Code Marketplace Install → VS Code — stable & Insiders

🔓 Open VSX Registry Install → Cursor, Windsurf, VSCodium, Theia, code-server, Gitpod, Antigravity, Kiro…

From inside the editor: open the Extensions view, search "OmniRoute" , click Install — works the same way on both stores. Source, issues and the publishing runbook live at diegosouzapw/OmniCopilot .

📖 VS Code Copilot Chat guide — setup, what the picker shows, dashboard-in-a-tab, troubleshooting

🔒 Private & Local-First

📖 Authorization · Guardrails · Compliance

🔌 Full CLI + A2A & MCP

Beyond the server, OmniRoute is a full command-line cockpit with 80+ commands , plus open agent protocols so an AI agent can drive it on its own .

⌨️ A real CLI (not just start )

omniroute # serve gateway + dashboard (port 20128) omniroute chat # interactive TUI chat client (slash: /model /combo /skill /memory) omniroute setup # guided first-run wizard omniroute doctor # diagnose providers, ports, native deps

🛰️ Remote mode — run the CLI here, OmniRoute on a VPS

OmniRoute on a server? Drive it from your laptop with the same CLI . Log in once with a scoped access token; every command then targets the remote.

omniroute connect 192.168.0.

15 # password → scoped token, saved as a context omniroute models list # ← runs against the REMOTE server omniroute configure codex # ← picks a remote model, writes a local Codex profile omniroute tokens create --name ci --scope read # mint narrower tokens for other machines omniroute contexts use default # ← switch back to the local server

Tokens are scoped read / write / admin ; process-spawning routes stay loopback-only. 📖 Remote Mode

🤝 Connect an agent — and it controls OmniRoute itself

Expose OmniRoute over MCP , A2A , a REST API , webhooks or a remote CLI — any capable agent (or your own code) gets the keys to the whole gateway: routing, providers, combos, cache, compression, memory — autonomously. HTTP endpoints below are served under http://localhost:20128 .

Interface Endpoint / command Use it for

🧰 MCP (stdio) omniroute --mcp Plug into Claude Desktop, Cursor, any MCP client

🌊 MCP (HTTP) /api/mcp/stream Remote MCP — 110 tools , 33 scopes (enforcement opt-in), full audit trail

📡 MCP (SSE) /api/mcp/sse Streaming MCP transport

🤝 A2A /.well-known/agent.json Agent-to-agent, JSON-RPC 2.0 + SSE, 6 skills

🌐 REST API /v1/* OpenAI-compatible — chat, embeddings, images, audio, OCR

🔔 Webhooks /api/webhooks Push request / quota events to Slack, Discord, Telegram or any URL

🛰️ Remote CLI omniroute connect Drive a remote instance with scoped access tokens

Give Claude Code the full OmniRoute toolset over MCP:

claude mcp add-server omniroute --type http --url http://localhost:20128/api/mcp/stream

📖 MCP Server · A2A Server · Agent Protocols

🗜️ Save 15–95% Tokens — Automatically

📖 How it works — pipeline, architecture & savings math

Default stacked combo runs RTK → Caveman . When both act on the same tool/context payload, savings compound:

combined = 1 − (1 − RTK) × (1 − Caveman_input) average = 1 − (1 − 0.80) × (1 − 0.46) = 89.2% range = 78.4 – 94.6%

Code blocks, URLs, JSON and structured data are always protected by the preservation engine.

Why use many tokens when few tokens do the trick? Every request passes through OmniRoute's compression pipeline transparently — no client changes.

It's now a stack of 12 composable engines that run in order and mix & match per routing combo — building on ideas from RTK , Caveman (⭐ 90K+), LLMLingua-2 , and Troglodita (PT-BR).

🧱 The 12-engine stack

Engines run in pipeline order; each is independently toggleable and configurable per combo:

Engine What it does

1 Session-Dedup Drops content repeated across turns (content-addressed, cross-turn)

2 CCR Archives large blocks behind retrieve markers, fetched on demand

3 Lite Whitespace + image-URL trimming (latency-light baseline)

4 RTK Smart tool-result filtering, dedup & truncation (command-aware)

5 Responses Tool Output Lossless-first JSON + bounded diagnostic compression for shell/patch/search/build outputs (Responses API)

6 Headroom Lossless tabular compaction of JSON arrays (~30%) via a vendored GCF codec

7 Relevance Extractive sentence scoring against the last user query

8 Caveman Rule-based prose compression (~65–75% on output)

9 Aggressive Summarization + progressive aging of old turns

10 LLMLingua-2 ML semantic pruning via MobileBERT ONNX — code-safe, async

11 Ultra Heuristic token pruning with an optional small-model (SLM) tier

12 OmniGlyph Experimental context-as-image encoding for measured Claude Fable 5 on the direct Anthropic wire; GPT 5.6 transformers remain fail-closed pending provider receipts.

Four compression profiles (aggressive default, balanced, coding-safe, passthrough) (most aggressive; opt-in)

Code blocks, URLs and structured data are always preserved byte-perfect. One-click presets combine the engines:

Mode Savings Best for

🪶 Lite ~15% Always-on safe default

🪨 Standard (Caveman) ~30% Daily coding

⚡ Aggressive ~50% Long tool-heavy sessions

🔥 Ultra ~75% Maximum savings

🧰 RTK 60–90% Shell/test/build/git output

🔗 Stacked (RTK → Caveman) 78–95% Mixed prompts + tool logs

Real example — Standard mode:

Before (69 tokens): "The reason your React component is re-rendering is likely because you're creating a new object reference on each render cycle.

When you pass an inline object as a prop, React's shallow comparison sees it as a different object every time, which triggers a re-render. I would recommend using useMemo to memoize the object."

After (19 tokens): "New object ref each render. Inline object prop = new ref = re-render. Wrap in useMemo."

Same answer. 72% fewer tokens. Zero accuracy loss. ✅

PT-BR example — Troglodita mode:

Antes (42 tokens): "O problema é que o componente está re-renderizando porque uma nova referência de objeto está sendo criada em cada ciclo de renderização. Eu recomendaria usar useMemo."

Depois (12 tokens): "Re-render: ref nova cada ciclo (objeto inline recriado). Usar useMemo ."

Mesma resposta. ~70% menos tokens. Precisão técnica intacta. ✅

🎚️ Beyond the engines — output styles, the adaptive dial & per-request control

The 12 engines above shrink what goes in . Three more layers shape how , when , and what comes out :

• 🪄 Output Styles (output-axis steering) — inject deterministic, cache-safe response-shaping instructions; combinable, each at lite / full / ultra intensity. Adding a style is a one-line registry entry:

• Terse prose — drop filler / articles / hedging; keep technical substance exact.

• Less code — "lazy senior dev" YAGNI: smallest working change, no unrequested scaffolding.

• Ponytail (lazy senior dev) — climb the YAGNI ladder, fix the root cause, smallest working diff.

• I have ADHD (action-first) — next action leads, steps numbered, one concrete next step, no preamble.

• Terse CJK (文言) — classical-Chinese ultra-terse style (locale-gated to zh ).

• 🎯 Adaptive context-budget (the dial) — instead of one on/off token threshold, escalate the cheapest, most-lossless engines only as far as needed

— 本文由 AI 根据公开来源辅助整理，命令、版本与许可证请在使用前到原始页面复核。