Blog
Long-form. Practical.
Cost optimization, fixing real errors, how kRouter works, comparisons, provider reviews and launch notes. Published by Kodelyth, the team that builds kRouter.
Xcode with any model: chat, Claude Agent and Codex
Xcode 26 and 27 can run chat, Claude Agent and Codex on other models, but each reads its own config. How to point all three at one local router.
Save money
· 17 postsGLM, MiniMax and Alibaba coding plans: stack two of them
What the GLM, MiniMax and Alibaba coding plans meter, where their 5-hour and weekly windows bite, and how to fall back from one plan to the next.
Qwen Code's free tier is gone: keep the CLI, swap the model
Qwen OAuth's free tier closed on April 15, 2026. What still works in Qwen Code, where Qwen models come from now, and how to keep the CLI answering.
Local first, cloud when it can't: Ollama with cloud fallback
Use a local Ollama model and switch to a cloud model only when it fails. What triggers the switch, what silently does not, and how to set it up.
Gemini CLI's free tier is gone: what still works for free
Google stopped serving Gemini CLI to free, AI Pro and Ultra users on June 18, 2026. Who kept access, what is still free, and how to keep using Gemini.
Caveman vs Ponytail: two ways to cut output tokens
Caveman cuts the words, Ponytail cuts the code. What each prompt does, what their benchmarks measured, and why you should run only one of them.
Free AI coding assistant: what is actually free in 2026
Free agents are easy; free models are the catch. What is still free in 2026 after Gemini CLI's free tier ended, what it costs you, and how to chain it.
Claude Code without a subscription: 6 backends that work
Six backends that can run Claude Code instead of a Claude plan: what each costs in 2026, which models it carries, and where each one breaks.
Use your GitHub Copilot plan in Claude Code and any tool
Copilot plans now include monthly AI Credits spent at API rates. How to spend them from Claude Code, scripts and other tools, and when that is worth it.
Codex CLI free: use your ChatGPT plan in every tool
Every ChatGPT plan includes Codex, yet your other tools still bill the OpenAI API. How to use the Codex models you already pay for in every tool.
Free Claude API access in 2026: the routes that work
Anthropic has no free API tier, but Claude runs on Kiro's and Antigravity's free plans and on plans you may already pay for. What each one limits.
ANTHROPIC_BASE_URL: point Claude Code at any model you want
Claude Code sends every request to ANTHROPIC_BASE_URL if the endpoint speaks Anthropic's Messages API. How to use it with any model, and the gotchas.
Headroom: the token saver that reclaims your context window
kRouter's Headroom saver cut one code-heavy request by 92.4%. It needs a Python proxy with the ml or code extras, is off by default, and fails open.
The true cost of Cursor Pro now that it bills at API rates
Cursor Pro is still $20 a month, but it no longer buys a request count. How the usage pool drains, what overage costs, and where your own keys fit.
Why your Cline bill hit $200 a month, and how to cut it
A Cline bill grows with every file it reads and every command it runs, because each step resends them. Where the tokens go and what really cuts them.
The cheapest LLMs for coding in 2026: real list prices
DeepSeek, GLM, MiniMax and Kimi priced from their own pages, flat-rate plans from $10, the free tiers still standing, and a combo that stacks them.
5 free Antigravity-powered coding setups for 2026
Antigravity's free plan serves Gemini 3.1 Pro and Claude under weekly limits. Five $0 coding stacks built around it with kRouter, and full setup steps.
Kiro AI vs Claude Pro: which is actually cheaper in 2026?
Kiro meters usage in credits; Claude Pro is a usage-capped plan with no API key. Current prices, how each runs out, and kRouter combos that use both.
Fix an error
· 17 postsScreenshots and text-only models like GLM-5.3: what works
GLM-5.3 can't read a pasted screenshot. Three fixes -- a model that reads images, a vision MCP server, a per-turn switch -- and where each falls short.
OpenRouter free tier limit: why you get 429s and what to do
OpenRouter's free models allow 20 requests a minute and 50 a day (1,000 after $10), and extra keys add nothing. Which 429 you hit and how to keep working.
Cursor ignores your localhost base URL: why, and the fix
Cursor calls a custom base URL from its own servers, so localhost never answers. How to expose a local router or Ollama safely, and the tunnel's limits.
Codex usage limit reached: keep working until it resets
Codex says you've hit your usage limit. What the five-hour and weekly caps count, when the message is wrong, and how to keep working until the reset.
Codex "stream disconnected before completion": every cause
Codex fails on a custom provider with "stream disconnected before completion". What the error means, its five causes, and how to tell them apart.
Claude Code weekly limit reached: keep working until reset
Claude Code says you've hit your weekly limit. What the caps count, what does not help, and three ways to keep working until the reset.
Claude Code context window with non-Claude models: the fix
Claude Code assumes 200K for a model it doesn't recognize, so it compacts too early or overruns the real limit. How to declare the real window.
Claude Code 529 Overloaded: who sent it and how to fail over
A 529 is not your quota: the model is out of capacity, and Claude Code has already retried. How to tell who sent it, and how to keep working.
OpenCode MissingSessionID: the x-opencode-session fix
OpenCode Go now returns HTTP 400 MissingSessionID for any request without an x-opencode-session header. What changed, why proxies broke, and the fix.
OpenAI insufficient_quota when you still have credits
insufficient_quota rarely means a zero balance. Usually billing was never activated, the credits expired, or the key's project has no budget.
Gemini 429 RESOURCE_EXHAUSTED: rate limit or quota?
Google returns the same 429 for a per-minute limit and a spent daily quota. How to tell them apart, when the daily quota resets, and how to fail over.
Cursor "You've hit your usage limit": three ways around it
Cursor's usage limit is two dollar pools, not a request count. Which one you spent, what on-demand costs, and how to keep working without upgrading.
Claude Code "rate limit exceeded": every cause and the fix
Claude Code stops with a 429 or a usage limit. Four different limits cause it, each needs a different fix, and the one everyone tries first makes it worse.
Antigravity account keeps getting banned: what helps
An Antigravity "verify your account" ban that keeps returning is a policy problem first. What Google's terms say, and how kRouter's backoff limits damage.
Duplicate AI replies? The response cache gotcha explained
Your AI answers a new question with its last reply? A response cache keyed on the wrong fields is a common cause. How it happens and how to check for it.
Anthropic rate limits explained: tiers, spend caps, 429s
Claude's API and Claude's plans limit you in different ways. How tiers, ITPM, spend caps and the 5-hour window work, and what actually helps.
Antigravity "verify your account" 403: what it means
Why Antigravity returns "Verify your account to continue", what Google's terms say about third-party tools, and what kRouter does with a flagged account.
How kRouter works
· 27 postsResponses API vs Chat Completions: why proxies break
Codex, some Copilot models and OpenCode Go's Grok speak only OpenAI's Responses API. Five ways it differs from Chat Completions, and where proxies break.
OpenCode Go in Claude Code and Codex: three protocols
OpenCode Go serves its models on three APIs. Which model needs which, why the wrong one fails with ModelProtocolUnsupported, and how to fix it.
JetBrains AI Assistant BYOK: any model through one endpoint
AI Assistant's OpenAI-compatible provider takes one URL. Put kRouter behind it for chat on any model, and see why Junie and completion stay out.
Copilot BYOK custom endpoint: any model in VS Code chat
VS Code's Custom Endpoint lets Copilot Chat use any Chat Completions, Responses or Messages API. Using it with a local router, and what still needs GitHub.
Codex CLI with any model: model_providers and wire_api
Codex now speaks only the Responses API. How to run it on Claude, GLM, Kimi or DeepSeek through config.toml, and what breaks along the way.
Kilo Code codebase indexing: picking an embedding provider
What Kilo's indexer sends, which embedding model to use for code, why changing it means a full re-index, and how to route it all through one endpoint.
Claude Cowork with other models: third-party inference setup
Claude Desktop can run Cowork through any Messages API gateway. Setting it up with kRouter, what the app assumes about your models, and what you give up.
Redact PII before it reaches the model, not after
kRouter can strip names, emails and keys from a request before it leaves your machine, and rejects the request if redaction fails instead of sending it.
What is an LLM gateway, and do you actually need one?
Gateway, proxy, router, aggregator: four words for overlapping things. What each one does, the problem they solve, and when not to run one.
kRouter is a multimodal AI gateway, not just an LLM router
One local endpoint for chat, speech, transcription, images, video, embeddings, web search and fetch across 95+ providers -- and where each one stops.
Live model catalog: new models without a kRouter update
kRouter asks each provider what it serves right now, so a model launched today can be added from the dashboard without waiting for a kRouter release.
Kiro API key (ksk_) auth: run Kiro on a headless server
Kiro API keys let kRouter use Kiro with no browser sign-in, on a VPS, in CI or Docker. They need a paid Kiro plan. Setup, curl import and caveats.
Use GitHub Copilot as a native Claude endpoint with kRouter
kRouter sends Claude requests on GitHub Copilot to Copilot's Anthropic-style /v1/messages endpoint, so prompt caching works. How it works, and its limits.
Claude Desktop, Antigravity and Copilot through a MITM proxy
Some AI apps have no base URL setting. How kRouter's MITM mode routes Claude Desktop, Antigravity, Copilot and Kiro anyway, and how to undo it cleanly.
One API for AI video: routing Grok Imagine through kRouter
kRouter proxies xAI's Grok Imagine video jobs on the same local endpoint as chat, audio and images. How submit-and-poll works, and what it does not do yet.
Kiro IDE with any model through a local MITM proxy
Kiro IDE has no custom endpoint setting. kRouter's MITM mode sends its chat turns to any model you map. How it works, the v0.5.137 fix, and setup.
Building a sub-5ms failover proxy: moving from SQLite to RAM
How kRouter took SQLite off the failover path: account locks are applied in memory first and written to disk in the background. What changed and why.
The Zenith routing engine: how kRouter picks an account
Zenith scores every account on success rate, latency and remaining quota before each request. The exact formula, what it fixes and what it does not.
kRouter Agent Skills: one link teaches an agent an endpoint
kRouter ships 8 Agent Skills: Markdown files an AI agent reads to call chat, TTS, STT, images, embeddings, web search and fetch on your own router.
How to use kRouter as a unified text-to-speech API
One /v1/audio/speech endpoint for 15 text-to-speech providers, from OpenAI and ElevenLabs to Edge TTS, three of them free with no key.
The security checklist for self-hosting an AI router
Self-hosting kRouter puts your OAuth tokens, API keys and prompts on that machine. What it protects by default, and what is still your job.
How to build a cheap autonomous coding agent with kRouter
A small TypeScript coding agent on kRouter's OpenAI-compatible endpoint, with free and cheap models behind one combo name, and what kRouter handles for it.
How to route Claude through AWS Bedrock with kRouter
kRouter has no built-in Bedrock provider, but Bedrock's Anthropic Messages endpoint works as a custom one. Setup, model IDs and the limits.
OpenAI to Claude to Gemini: how kRouter translates formats
Every AI provider speaks its own JSON dialect. How kRouter converts OpenAI, Claude, Gemini, Responses and Kiro requests live, with before/after JSON.
Self-host an AI router on a $6/month VPS
Deploy kRouter on a small VPS so your team shares one endpoint: systemd, nginx with streaming, Tailscale-only access, Docker, and the caveats.
Multiple accounts per provider: how kRouter rotates them
Connect several accounts of one provider and kRouter picks one per request: Zenith scoring, round-robin, cooldowns, concurrency caps, and the terms risk.
RTK token compression: what it trims from tool output
RTK, on by default in kRouter, trims diffs, build logs, listings and long file reads in tool results before the request leaves your machine.
Comparisons
· 14 postsGrok in Claude Code and Codex: three ways to pay for it
Grok 4.5 to 4.7 in Claude Code or Codex through Grok Build, OpenCode Go or the xAI API: what each costs, which API it speaks, and the setup.
Claude Code with GLM, Kimi or DeepSeek: setup and trade-offs
GLM-5.3, Kimi K3 and DeepSeek V4 can all drive Claude Code. What each costs, which slot each suits, what breaks, and when you need a router.
Best self-hosted LLM gateways in 2026, and how to vet one
LiteLLM, Bifrost, Portkey, LLM Gateway, Agent Router and kRouter: who each is for, and the checks to run before a gateway holds your API keys.
Best AI coding assistant in 2026: how to actually choose
Three questions decide it, not a feature table. Cursor, Cline, Continue, Claude Code, Codex, Kiro, Antigravity and Gemini CLI on what actually differs.
OpenRouter alternative: when self-hosting is worth it
OpenRouter is one key, hundreds of models, zero ops and a fee on every credit top-up. When self-hosting beats it, and when it honestly does not.
LiteLLM vs kRouter: they solve different problems
LiteLLM is a Python gateway for teams that need budgets and spend tracking. kRouter is a local router for one developer's subscriptions. Pick by audience.
GitHub Copilot vs Cursor: you may already be paying for both
Copilot and Cursor now both bill model use at API rates from a monthly pool, and their model lists overlap heavily. What each plan buys beyond the models.
Gemini CLI vs Claude Code: what changed in 2026
Gemini CLI lost its free tier in June, and both tools now run 1M-token models. What still separates them, and how to use Gemini models inside Claude Code.
Cline vs Cursor: the bill is the whole difference
Cursor includes a set amount of model usage in a subscription; Cline is free and bills per token. Since Cursor went usage-based, the gap is narrower.
Claude Code vs Codex CLI: same job, different guardrails
Both read your repo, edit files and run commands. They differ in how they act without asking, what they cost, and how you point them at other models.
Cursor alternatives that are actually free in 2026
Free, open-source editors that do Cursor's job -- Cline, Kilo Code, Aider -- and where to get the model for each now that several free tiers have closed.
Windsurf vs Cursor in 2026: Windsurf is now Devin Desktop
Windsurf became Devin Desktop in June 2026 and both editors now meter usage at API rates. How the two compare on plans, limits and your own keys.
Ollama vs LM Studio: the local AI showdown for 2026
Ollama, LM Studio, llama.cpp and vLLM all run open-weight models locally. What each does well in 2026, what changed, and when to add cloud fallback.
Claude Code vs Cursor in 2026: agent, editor, and the bill
Claude Code is an agent that lives in your terminal; Cursor is an editor with a usage pool. What actually differs in 2026, what each costs, and using both.
Provider reviews
· 2 postsFeatherless, Venice AI and Perplexity Agent in kRouter
Featherless for open models, Venice for private inference, Perplexity Agent for many labs on one key: what each offers and how to connect it to kRouter.
DeepSeek V4 vs GPT-5.5 and GPT-6 for coding: what it costs
DeepSeek V4 Pro and Flash against GPT-5.5 and GPT-6: official per-token prices, DeepSeek's peak-hour rates, and how to run both behind one endpoint.
