cortiq-dsh-llm-router 🧠✨
Your LLMs on autopilot. One key — every prompt is automatically dispatched to the best model.
Stop guessing which model to pick. Just type — the plugin figures out where to send each request.
🚀 Quick Start (1 minute)
dsh plugin add cortiq-dsh-llm-router
# Go to Settings → Models → Cortiq Router, paste your key
# Then pick "Cortiq Router · Auto" in the chat model selector
Done. The plugin handles everything from here.
One prerequisite: the router picks among the models your other provider
plugins already serve (e.g. llm-deepseek). Keep at least one of them
configured — the router never talks to a model vendor itself.
🎁 Free Trial
Get a free key for 1 month — no payment required.
👉 allaigate.com — sign up and get a cortiq_… key instantly.
No promo code needed. After the trial, plans start at just $1/month.
💰 What does it cost?
| Without plugin | With plugin |
|---|
| All requests go to expensive models | Simple requests → cheap model |
| Overpay 10–50× | Save up to 90% |
| Manual model selection every time | Fully automatic |
🎯 What it does
The plugin is a smart dispatcher for your LLMs. It reads every prompt, understands what you're asking, and routes it to the right model.
Live example
You type: "Write a Python function to reverse a linked list"
| Without plugin | With plugin |
|---|
| 🤷♂️ Have to pick a model yourself | 🔍 Router: "This is code, low complexity (0.29)" |
| 🎲 Guess: DeepSeek V4 Pro ($) | 🧠 Rule: code @ low → fast model |
| Response in 3s, paid $0.01 | ⚡ Response in 0.5s, free |
But: "Implement a distributed consensus algorithm"
| Without plugin | With plugin |
|---|
| 🤷♂️ Guess again | 🔍 Router: "This is code, high complexity (0.78)" |
| Maybe Flash is enough? | 🚀 Escalation: hard request → powerful model |
| Response might be bad | ✅ Strong model handles it well |
⚙️ What can you configure?
Complexity thresholds
complexityBands:
low: 0.35 # score ≤ 0.35 → simple task
medium: 0.65 # ≤ 0.65 → medium, > 0.65 → hard
Which models at which complexity
globalTiers:
low: [deepseek-v4-flash] # simple → fast & cheap
medium: [deepseek-v4-pro] # medium → more powerful
high: [deepseek-v4-pro] # hard → strongest
Per-task rules
taskRules:
code:
low: [deepseek-v3-chat] # simple code → fast model
high: [deepseek-v4-pro] # complex code → powerful model
translation: [deepseek-v3-chat] # simple form: same model at all tiers
creative-writing:
low: [deepseek-v4-flash]
high: [deepseek-v4-pro]
🎮 Who is this for?
| You | Why you need it |
|---|
| AI agent developers | Your agent picks the right model automatically, saving tokens |
| Freelancers / solo devs | Save on API costs without sacrificing quality |
| Teams / startups | One key, automatic routing, full transparency |
| Enthusiasts | Smart routing for $1/month — no headache |
📊 Real-world accuracy
We ran 8 different prompt types through the live allaigate router:
Prompt Task Conf → Routed To
─────────────────────────────────────────────────────────────────────────────────────────────
Write a Python function to reverse a linked list code 0.991 fast model ✓
Solve quadratic equation x² - 5x + 6 = 0 math 0.980 medium model
Translate to French: Hello world translation 0.988 fast model ✓
Write a poem about AI consciousness creative-writing 0.950 fast model ✓
What is the capital of France? qa 0.998 medium model
Summarize: The quick brown fox... summarization 0.980 medium model
Extract date and price from invoice extraction 0.997 medium model
How are you doing today? chitchat 1.000 medium model
8/8 — all task types correctly identified.
✓ = task rule applied → model chosen intentionally
🔄 How it works
You: "write code to reverse a list"
│
▼
┌─────────────────────────────┐
│ 1. Text extraction │ ← only last_user (PII stays local)
└──────────┬──────────────────┘
│ POST /v1/route
▼
┌─────────────────────────────┐
│ 2. Semantic router │ ← router.allaigate.com
│ task=code, complexity=low │
└──────────┬──────────────────┘
│
▼
┌─────────────────────────────┐
│ 3. Routing table │ ← your configuration
│ code @ low → flash │
│ ↳ fallback → flash/pro │
└──────────┬──────────────────┘
│ invoke
▼
┌─────────────────────────────┐
│ 4. Delegated to the provider │
│ plugin that serves it │
└─────────────────────────────┘
🎛 Picking a policy per chat
The model picker lists four entries. All of them route; the last three pin the
policy profile for that conversation, so you don't have to open settings:
| Entry | What it does |
|---|
| Auto | Routes using the profile configured in settings |
| Auto · cost saver | Prefers the cheap end of every chain |
| Auto · balanced | The middle ground |
| Auto · quality first | Escalates sooner to the strong models |
⚙️ Settings UI
The plugin adds its own Cortiq Router section to the harness settings, in
every language it ships:
| What you get there | |
|---|
| Classifier key | paste it once; stored through the credentials service, with a configured / not-configured badge |
| Endpoint & behaviour | classifier URL, policy profile, what gets classified, character cap, timeout, decision logging |
| Complexity bands | where low / medium / high sit on the 0…1 score, with a low-below-medium check |
| Model chains | the per-tier chains and the fallback model |
| Per-task rules | a table of task × low / medium / high / any, with add and remove |
It is a section of its own, not a card inside Models. The Models page
picks its provider editor by namespace and offers third-party providers only
a note pointing at settings.yaml, with Apply disabled — so the plugin
registers through the same settings.section slot the Models page itself uses.
🌍 Languages
The plugin ships its own copy in 7 languages — 🇬🇧 English · 🇷🇺 Русский ·
🇨🇳 中文 · 🇩🇪 Deutsch · 🇫🇷 Français · 🇪🇸 Español · 🇹🇷 Türkçe — selected by the
language setting:
llm-cortiq-router:
language: auto # auto | en | ru | zh | de | fr | es | tr
auto follows the harness UI language. Note that the dsh browser client itself
ships only English and Chinese, so auto resolves to one of those two; pick any
of the other five explicitly here to use it.
Translated: the entries this provider shows in the model picker, and the message
shown when no candidate model could serve a call — the strings a provider plugin
actually puts in front of you. Log lines stay English: they are diagnostics meant
for issues and grep.
📦 Installation
# From DSH plugin store
dsh plugin add cortiq-dsh-llm-router
# Or via npm
npm install cortiq-dsh-llm-router
Or add to your cordis.patch.yml:
- insert:
- id: llm-cortiq-router
name: 'cortiq-dsh-llm-router'
🔑 Get Your Key
- Open allaigate.com
- Sign up — get a free 1-month key instantly
- Copy your key (
cortiq_…)
- In DSH: Settings → Models → Cortiq Router → paste your key
Or set the environment variable:
export CORTIQ_ROUTER_KEY=cortiq_your_key_here
💡 Tips
- Start with defaults — they're sensible and work out of the box; the
classifier lives at
router.allaigate.com, and routerUrl points a
self-hosted deployment somewhere else
- Add per-task rules for tasks that matter most to you
- Tweak complexityBands to fine-tune when a request is considered "hard"
- Watch the log — with
echoRouting on (the default), every call prints
task=… tier=… score=… → provider/model, so you can see what it chose
- Chains survive a missing key — a candidate whose provider refuses the
call (no credential, no adapter, unknown model) is skipped and the next one
answers; only a provider that already started streaming keeps the call
- Name a route explicitly — a candidate written as
provider:model
(e.g. deepseek-official:deepseek-v4-pro) skips catalog lookup entirely
📄 License
MIT — do whatever you want.
One key. No guessing. Every request goes to the best model. 🚀
