The AI coding agent router

The right intelligence for every step.

AirLLM understands what your coding agent is doing — planning, editing, running tools, checking results or recovering from failure — and routes each step to the intelligence it actually needs.

Strong models when they matter. Fast models when they don’t.

Start Free Connect Your Tools

Free plan · No card required

Works with

  • Claude Code
  • Codex CLI
  • Cursor
  • GitHub Copilot CLI
  • OpenCode
  • Aider
  • AirCode

Keep your coding tool. Change the intelligence behind it.

Products

Pick How You Want to Code

Three ways onto the same platform — the AirCode agent as a native macOS app or in your terminal, or the coding tool you already use through Air Router. One key, one bill, the same savings.

Vibe coding

AirCode Studio

Build, run and verify Web, iOS and Android apps with the AirCode agent in a native macOS workspace.

Explore Studio
Built for pros

AirCode CLI

TUI-based development in any terminal.

Install CLI
Keep your agent

Air Router

Continue in Claude Code, Codex, Cursor, or Copilot through Air Router and get the savings without changing your workflow.

Connect Router

A coding session

One coding task. Different intelligence at every step.

A coding task isn’t one step. Air Router treats every step differently.

UserAdd OAuth login and update the tests.
  1. 01Understand repositoryFast modelRoutine
  2. 02Design authentication flowStrong modelPlanning
  3. 03Read configurationFast modelRoutine
  4. 04Modify authentication serviceStrong modelCode edit
  5. 05Run testsFast modelTool step
  6. 06Diagnose failing testStrong modelReasoning
  7. 07Verify implementationStrong modelVerification

One task. Seven steps. No reason to pay for maximum intelligence seven times.

Air Router

Most routers choose a model. Air Router manages the journey.

Traditional routing asks: Which model should answer this prompt?

Air Router asks: What does the agent need at this point in the task?

Prompt router
  1. Prompt
  2. Classify
  3. Choose model
  4. Answer
Air Router Agent loop
  1. Agent session
  2. Understand step
  3. Check task progress
  4. Check cache economics
  5. Choose intelligence
  6. Evaluate result
  7. Escalate · Continue · Stop

Routing becomes part of the agent loop — not just the API call.

01 · Escalate

Intelligence only moves one direction.

AirLLM can start efficiently — but once a task proves difficult, it doesn’t blindly downgrade the session again.

Up.

Task intelligenceNo step back
  1. Fast
  2. Capable
  3. Strong
  4. Strongest

No downgrade mid-task

Once the task requires stronger intelligence, AirLLM is designed not to drop it back down halfway through the work.

Resend to upgrade

Send the same request again and AirLLM can automatically try a stronger tier.

Weak answer? Escalate.

Rejected or broken responses can be retried on stronger intelligence instead of simply being passed back to the agent.

Start efficient. Escalate when evidence says you should.

02 · Optimize

The lowest-priced model isn’t always the cheapest request.

Coding agents repeatedly send large amounts of context. Switching models can destroy a valuable warm prompt cache.

AirLLM evaluates the economics of the actual request — not simply the model’s advertised input-token price.

Option A

Switch model

Token priceCheap

  • Rebuild context
  • Cold cache
  • Another agent turn

Option B

Stay

Token priceHigher

  • Warm cache
  • Cached context
  • Fewer repeated tokens
Air Router’s pick

Illustrative comparison — the effective cost depends on the context size and how warm the cache is.

Cache-aware switching

Don’t abandon a warm model unless switching still makes economic sense.

Tokenizer-aware cost

Compare models using how providers actually count and bill the request.

Session affinity

Keep useful provider and model caches warm across the coding session.

How It Works

How AirLLM Saves You Money

Same workflow. Smarter backend. Lower token cost.

  1. Your Tool Sends a Request

    AirCode CLI, Claude Code, Codex, Copilot, Cursor, agents, or your app sends a normal AI request.

    • AirCode CLI
    • Claude Code
    • Codex
    • Copilot
    • Cursor
    • AI Agent

  2. AirLLM Optimizes the Input

    AirLLM analyzes the prompt, removes waste, applies caching, and checks reusable context.

    Up to 75% token reduction

  3. AirLLM Classifies the Task

    The request is classified by difficulty, task type, required quality, and expected cost.

    • Coding
    • Reasoning
    • Simple Q&A

  4. AirLLM Routes to the Right Model

    The request is sent to the best model for that job, balancing quality, speed, and cost.

    A right-sized model for each request

  5. You Get the Response Back

    AirLLM returns the final response to your tool in the expected format. Zero friction.

    Drop-in compatible, no changes

Same workflow. Smarter backend. Lower token cost.

No code changes. No team retraining. AirLLM is invisible by design.

See How It Works Live

03 · Protect

When your agent gets stuck, your bill shouldn’t keep moving.

AirLLM watches for repetitive agent behavior, attempts to push the agent toward a different approach, and can stop a runaway session before it continues wasting tokens.

Sometimes the smartest routing decision is to stop routing.

agent session
1Read auth.ts
Edit auth.ts
Run tests — FAIL
2Read auth.ts
Edit auth.ts
Run tests — FAIL
3Read auth.ts
Edit auth.ts
Run tests — FAIL
Nudge agent→Still stuck?→Stop session

Built for long-running agents

Your agent shouldn’t care which provider failed.

Automatic failover

Rate limits, provider errors and timeouts can move requests to another route.

Bad starts intercepted

Failures from new or uncertain routes can be caught before reaching the client.

No empty answers

Empty model responses aren’t treated as successful work.

Duplicate requests collapsed

Identical concurrent requests can share one upstream execution.

AirCode

What happens when the agent and router understand each other?

AirCode is the coding agent built for AirLLM.

Instead of forcing the router to infer everything from a generic API request, AirCode can expose richer information about what the agent is doing — creating a tighter loop between execution, routing and verification.

Explore AirCode Install AirCode CLI

AirCode Loop

Give it a goal. AirCode works toward passing checks.

AirCode Loop can work toward an objective across multiple rounds while staying inside explicit limits. Each round can be checkpointed so regressions don’t become the starting point for the next attempt.

Round limitTime limitSpend limit

Every round

  1. Checkpoint createdthe known-good state
  2. Change attemptedthe round’s edit
  3. Tests regressworse than the checkpoint
  4. Rollbackback to the checkpoint

Rollback is there so regressions don’t become the starting point for the next attempt.

Autonomy with an exit condition.

AirCode verification

“Done” shouldn’t mean the model stopped typing.

AirCode turns the task into an acceptance checklist and runs the project’s checks before declaring the work complete. You stay in control: review its changes before you ship them.

Task

Add OAuth login

Acceptance

  • Login endpoint implemented
  • Existing authentication preserved
  • Error path handled
  • Tests updated

Project checks

  • npm test
  • build
  • lint
VERIFIED

Completion is a testable state, not a model opinion.

AirCode Watch

Your code changes. AirCode reacts when something breaks.

AirCode Watch can run your project’s checks when files change and involve the model only when something actually fails.

Save file Run checks
Pass→Do nothing — no model call
Fail→AirCode investigates→Fix→Run checks again
watch · on saveNo failure. No model call.

One system

One intelligence layer. Three ways to use it.

The platform

AirLLM

Unified access, billing, usage, governance and infrastructure for AI models.

Explore Platform
The intelligence layer

Air Router

Understands tasks, routes agent steps, protects cache economics, escalates quality and helps control wasted work.

Explore Air Router
The coding agent

AirCode

A coding agent built to take full advantage of Air Router — with verified loops, checkpoints, watch mode and deep session continuity.

Explore AirCode
AirCodeAir RouterAirLLMClaude · OpenAI · AirGPU · Other providers

Use the whole stack — or connect the coding tools you already have.

Your tools

You don’t have to switch coding agents.

Air Router works behind the tools developers already use.

  • Claude Code
  • Codex CLI
  • Cursor
  • GitHub Copilot CLI
  • OpenCode
  • Aider

Want the agent built specifically for the router? Use AirCode

Model families

Route intelligently without leaving your preferred model family.

airllm-anthropic

AirLLM chooses the appropriate Claude tier for the task while keeping the route inside the Claude family.

Haiku→Sonnet→Opus
airllm-openai

Stay inside the OpenAI family while AirLLM adjusts intelligence to the task.

Lighter→Balanced→Strongest

Automatic routing doesn’t have to mean giving up control.

Security

Code goes through the router. Secrets shouldn’t.

Secret scrubbing

API keys, passwords and tokens detected in coding-agent traffic can be removed before model dispatch.

Sensitive-file protection

Files such as .env, keys and certificates can be prevented from being sent to models.

Hashed API keys

Platform API keys are stored as hashes and displayed once.

Organization controls

Administrators can restrict model routes across the organization.

Dashboard

See what your agent actually used.

Understand where intelligence was spent instead of receiving one mysterious monthly model bill.

Pricing

Start free. Scale when it proves useful.

Free$0
Student$10 / month
Pro$20 / month
Pro Plus$40 / month
Max$100 / month
Max Plus$200 / month
Pay as you goUsage-based

AirLLM doesn’t promise that every token is cheaper.

The goal is to reduce the cost of completing the work by using expensive intelligence selectively, preserving useful caches, reducing unnecessary context and stopping wasted agent turns.

View Pricing

FAQ

Questions, answered plainly.

Is AirLLM just another model router?

No. Traditional routers primarily choose a model for a request.

AirLLM is designed around coding-agent sessions: understanding different steps, maintaining task progression, accounting for cache economics, escalating weak results and helping control runaway behavior.

Is AirLLM cheaper per token?

Not necessarily. AirLLM doesn’t promise a lower price for every token.

AirLLM is designed to reduce the cost of completing a task by avoiding unnecessary premium-model turns, preserving useful prompt caches, reducing redundant context and stopping wasted work.

Will routing to smaller models reduce code quality?

Planning and difficult edits can remain on stronger models.

When a task proves more difficult, AirLLM can increase the intelligence tier without automatically downgrading it again during that task. Weak results can also be escalated.

Can I stay entirely on Claude?

Yes. Use airllm-anthropic to route between appropriate Claude tiers without leaving the Claude model family.

Use airllm-openai for the corresponding OpenAI-family experience.

Do I need AirCode?

No. AirLLM works with supported coding tools including Claude Code, Codex CLI, Cursor, GitHub Copilot CLI, OpenCode and Aider.

AirCode provides the deepest integration because the agent and router are designed together.

What happens when my plan allowance runs out?

Users can add prepaid credit or enable auto-recharge.

Without additional credit, paid-plan requests stop until the allowance resets. Free users can continue through available free-model routes.

Your coding agent already knows how to code. Give it smarter infrastructure.

Connect Claude Code, Codex, Cursor or AirCode and let AirLLM decide where each step should run.

Start Free Connect Your Tools

No card required. Setup in minutes.

Keep exploring

Related guides and products

Explainer

What is an AI coding agent router?

Why coding agents need per-step model routing, and what to look for in a router.

Read Product

Coding tools

Keep Claude Code, Codex, Cursor or Copilot CLI, and route them through Air Router.

Read Product

AI agent cost optimization

How Air Router optimizes the model requests your agents make.

Read Product

Routing models

airllm-auto, airllm-anthropic, airllm-openai and more: automatic model selection.

Read Knowledge

Guides for developers

Coding agents, LLM routing, cost optimization and AI infrastructure.

Read Comparisons

How AirLLM compares

Neutral comparisons with OpenRouter, LiteLLM, Portkey, Claude Code, Codex and Cursor.

Read