As of October 8, 2026, two new sections of Relaykit are live: Agent Kits, with setups for the coding agents developers use most, and Free Tools, three small utilities that run in your browser. Both are free and need no account. This post explains what each one does and why we built it.
Agent Kits: one page per coding agent
Every coding agent has its own way to set a custom endpoint. Claude Code reads an environment variable. Codex CLI uses a provider block in its config file. Cursor has a field in its settings screen. Cline and Kilo Code have an “OpenAI Compatible” provider option. The ideas are the same everywhere: a base URL, a key and a model ID. But the names, the file locations and the mistakes people make are different in each tool.
So each agent gets its own kit:
| Agent | Type | API format it uses |
|---|---|---|
| Claude Code | CLI | Anthropic |
| Codex CLI | CLI | OpenAI |
| Cursor | Editor | OpenAI |
| Cline | VS Code extension | OpenAI |
| Kilo Code | VS Code extension | OpenAI |
| OpenCode | CLI | OpenAI |
| Aider | CLI | OpenAI |
| Zed | Editor | OpenAI |
Each kit follows the same structure, so once you have set up one agent, the next is familiar:
- What you need. A Relaykit key and the agent installed. We link to the agent’s official install instructions. We don’t copy them, because they change.
- The configuration, ready to copy, with the right base URL for the agent’s format. Anthropic-format tools use
{SITE.apiBase}. OpenAI-format tools use{SITE.apiBase}/v1. - A check that it works, using the agent’s own status command or a one-line test prompt, so you know the agent reaches the endpoint before you start real work.
- Which model to pick, with our rates for that model.
- Troubleshooting for the errors people actually hit with that tool: a missing
/v1, an old key still set somewhere in the shell, or a model ID the tool rewrites.
We only cover tools that are actively maintained. When an agent changes how its configuration works, we update the kit and its date. When an agent is discontinued, we remove its kit rather than leave instructions that no longer work.
Which model for which agent
The kits suggest a default, but you choose. Some starting points, with our current rates per 1M tokens (input / output):
- Long agent sessions and large refactors: Claude Opus 5.5 at $4.40 / $22. Use it when the task needs careful multi-step work.
- Everyday coding: Claude Sonnet 5 at $2.20 / $11. Most sessions don’t need more.
- Sub-agents and quick edits: Claude Haiku 4.5 at $1.10 / $5.50.
- Codex CLI and OpenAI-format tools: GPT-6.1 Sol at $2.20 / $11.
- Budget batch work: DeepSeek V4 Flash at $0.24 / $0.96.
Our post on choosing a model for coding agents explains these trade-offs, and the models page lists all 11 models we serve.
One practical tip: give each developer’s agent its own key with a spend limit. Agent usage is uneven, and a separate key per person means one long session can’t use up the team’s balance. The per-project keys tutorial shows how.
Free Tools: three utilities that run in your browser
We kept wanting quick answers while building with LLM APIs: what will this cost per month, what exactly goes in this agent’s config file, roughly how many tokens is this prompt. The tools answer those questions. They run in your browser, and you don’t need an account.
Cost calculator
Choose a model, then enter requests per day and the typical input and output tokens per request. You get a monthly estimate, with input and output shown separately so you can see which one drives the cost. The rates come from the same catalogue as our models page, so the calculator and the price list always match.
It’s an estimate, and it’s only as good as your token numbers. The best inputs are averages from the usage objects of real responses. If you don’t have those yet, our guide to estimating monthly LLM API cost shows how to get reasonable first numbers.
Config generator
Choose an agent and a model, and you get the exact configuration for that tool: environment variables, a settings file or the values for a settings screen, with the right base URL for the agent’s format. Copy it to the place the config names, then add your key.
The generated config uses a placeholder for the key. Paste your real key on your own machine, into the file or variable the config names. Keys should go straight from the console to where they’re used, not through other websites, including ours.
Token estimator
Paste a prompt, a file or a conversation, and you get an approximate token count and what it would cost as input on each model.
To be clear about what “approximate” means here: each model family uses its own tokenizer, so the same text is a different number of tokens on Claude, GPT or Grok. The estimator uses a general approximation instead of each vendor’s tokenizer. It’s close enough for budgeting and for checking whether something fits in a context window. It is not an exact count. The exact count for any request is in that response’s usage object and in your usage log.
Why these are free
They’re useful, and they help developers find Relaykit. That is the whole reason. There is no sign-up wall and no trial timer. If one of them saves you time and you later need an API, we hope you’ll try ours. If you don’t, that’s fine too.
Feedback
If a kit doesn’t work with the current version of an agent, or a tool gives a number that looks wrong, email support@relaykit.dev with the agent’s version or the inputs you used. We fix kits when the tools change, and we post notable changes here in Product Updates.