Coding agents
Coding tools
AI coding agents burn the most tokens of any workload — and hit provider rate walls and subscription time windows the hardest. Every tool below has a native endpoint on BoostRail — Claude Code over Anthropic's Messages protocol, the rest over OpenAI — so switching to BoostRail takes about 30 seconds: point the base URL at BoostRail and drop in one key.
From zero to your first call
Create an API key
Sign up and create a key in the console. One key works across every tool on this page, on any model in the catalog.
Paste the config
Pick your tool below and copy its block. You change the base URL and the key — your prompts, workflow, and editor setup stay exactly as they are.
Confirm it works
Make one call before you start a long session. It takes a few seconds and tells you immediately whether the config landed.
Copy-paste config, one per tool
Claude Code speaks Anthropic's native Messages protocol against BoostRail's /v1/messages endpoint. The rest speak OpenAI — Codex CLI over /v1/responses, the others over chat completions. Either way it is the same key and the same catalog: set the base URL, paste your BoostRail API key, and the model id is any model from the catalog.
Anthropic-native protocol over /v1/messages. Put this in ~/.claude/settings.json so desktop launches and background agents pick it up too — shell exports alone don't reach them.
// ~/.claude/settings.json
{
"env": {
"ANTHROPIC_BASE_URL": "https://api.boostrail.com",
"ANTHROPIC_API_KEY": "YOUR_API_KEY",
"ANTHROPIC_MODEL": "claude-sonnet-5",
"CLAUDE_CODE_ATTRIBUTION_HEADER": "0",
"CLAUDE_CODE_DISABLE_EXPERIMENTAL_BETAS": "1",
"CLAUDE_CODE_ENABLE_GATEWAY_MODEL_DISCOVERY": "1"
}
}
// Verify: run claude, then /status —
// the base URL should show api.boostrail.com.
// Non-Claude models: set ANTHROPIC_MODEL to any catalog id,
// e.g. deepseek-v4.1-flash — the /model picker only lists claude ids.Add a provider block to ~/.codex/config.toml and export your key. Codex speaks the Responses protocol; BoostRail's /v1/responses endpoint serves it natively.
# ~/.codex/config.toml
model = "claude-sonnet-5"
model_provider = "boostrail"
[model_providers.boostrail]
name = "BoostRail"
base_url = "https://api.boostrail.com/v1"
env_key = "BOOSTRAIL_API_KEY"
wire_api = "responses"
# shell: export BOOSTRAIL_API_KEY=YOUR_API_KEYIn the Cline settings panel (VS Code), pick the OpenAI-Compatible provider and fill these fields.
# Cline → Settings → API Provider
API Provider: OpenAI Compatible
Base URL: https://api.boostrail.com/v1
API Key: YOUR_API_KEY
Model ID: claude-sonnet-5Declare BoostRail as a provider in opencode.json (project root or ~/.config/opencode/).
{
"$schema": "https://opencode.ai/config.json",
"provider": {
"boostrail": {
"npm": "@ai-sdk/openai-compatible",
"name": "BoostRail",
"options": {
"baseURL": "https://api.boostrail.com/v1",
"apiKey": "{env:BOOSTRAIL_API_KEY}"
},
"models": {
"claude-sonnet-5": { "name": "Claude Sonnet 5" }
}
}
}
}Add a model entry to ~/.continue/config.yaml with provider: openai and an apiBase override.
# ~/.continue/config.yaml
models:
- name: BoostRail Claude Sonnet 5
provider: openai
model: claude-sonnet-5
apiBase: https://api.boostrail.com/v1
apiKey: YOUR_API_KEYAider reads the standard OpenAI-compatible env vars; point them at BoostRail.
export OPENAI_API_BASE=https://api.boostrail.com/v1
export OPENAI_API_KEY=YOUR_API_KEY
aider --model openai/claude-sonnet-5- The snippets default to claude-sonnet-5 — the flagship most coding agents reach for. Swap the model id for any model in the catalog — claude-opus-4.8, gpt-5.4, deepseek-v4.1-flash, and more — without touching anything else. Same base URL, same key.
- YOUR_API_KEY is a placeholder: create a key in the console and paste it in.
- Running OpenClaw instead? Long-running agents have their own setup page, with the provider block for openclaw.json. OpenClaw setup
Confirm it is working
Run this once after pasting the config. It separates a configuration problem from a model problem, which is the difference between a two-minute fix and an hour of guessing.
One call, straight to the API
curl https://api.boostrail.com/v1/chat/completions \
-H "Authorization: Bearer YOUR_API_KEY" \
-H "Content-Type: application/json" \
-d '{"model":"claude-sonnet-5","messages":[{"role":"user","content":"reply with OK"}]}'A working setup returns a JSON body with a choices array and the model's reply inside it. Your usage and the exact price applied to that request appear in the console within a few seconds.
401 unauthorized
The key did not reach the request. Check that you pasted the key value itself rather than its name, and that the tool was restarted after the config change.
400 model_not_found
The model id is not in the catalog. Copy an id from the Models page — ids are exact strings, for example claude-sonnet-5.
curl works, the tool does not
The tool is still reading its old configuration. Claude Code needs the entry in ~/.claude/settings.json (shell exports do not reach desktop launches or background agents), and VS Code extensions need a window reload.
Why route your coding agent through BoostRail
Coding agents are the heaviest token consumers in the industry: long files, tool loops, and repeated context make agent load the norm, not the exception. On official subscriptions that means hitting a ~5-hour usage window; on raw API tiers it means bumping into per-minute rate limits mid-task.
BoostRail is one API across leading models, in the OpenAI and Anthropic formats. The tools you already use keep working unchanged — you only swap the base URL and key. No subscription window and no tier rate wall.
What you actually pay — and don't
The numbers that matter for a token-heavy coding workload:
Better price on agent load
Leading models are priced at or below their official list prices, and every model's price is public on the Models page. The live per-request price in the console prevails.
No usage window
No ~5-hour subscription window standing between the agent and the model. Long sessions and heavy loops run without a subscription time limit.
Coding tools FAQ
Why do some Qwen models carry a context cap?
Alibaba Cloud prices those models in usage tiers — longer inputs cost several times more per token. BoostRail sells each of them at one flat rate matching the official price of the covered tier, and caps input at that tier (≤120K for the coder pair, ≤240K for qwen3.7-plus/flash). Requests over the cap return context_length_exceeded before any charge; coding agents handle that by compacting automatically. For 1M-context work, qwen3.7-max has a single official rate and no cap.
Do I need a different key per tool?
No. One BoostRail API key works across every tool here — Claude Code, Codex CLI, Cline, OpenCode, Continue, Aider — against the same base URL.
Does Claude Code's extended thinking work?
Yes. Claude Code's thinking setting is passed on and translated for the route serving the model, so the model does its reasoning pass. The reasoning text itself is not returned as thinking blocks: you get the answer and the tool calls. If a route cannot apply the setting, the x-boostrail-degraded response header says so.
Anything else to know when using Claude Code with BoostRail?
Two details. Token counting: BoostRail does not serve /v1/messages/count_tokens, so Claude Code falls back to its own local estimation — that is the officially supported path. Images inside tool results (e.g. reading a screenshot file) are replaced with a text placeholder; regular image inputs in your messages work normally.
Does Codex CLI really work, given it only speaks the Responses protocol?
Yes. BoostRail serves /v1/responses statelessly: messages, function tools and streaming run on the same routing and billing as chat completions. Details are in the Responses reference under Docs.
Will switching break my existing setup?
No. You change only the base URL and the key; the tool, your prompts, and your workflow stay the same. You can point the base URL back to a provider at any time — zero lock-in.
How is agent load priced?
Per token in USD; leading models are priced at or below their official list prices. Where a model's route supports prompt caching, cached input bills at the cache-read price published for that model — at or below the provider's own cache-read price, DeepSeek V4 Pro, DeepSeek V4.1 Flash and Kimi K3 excepted — and cache writes carry no surcharge. The Models page publishes the live per-token price for every model.
Point your agent at BoostRail
Create a key, paste it into the tool of your choice, and keep coding — on any model in the catalog.
Last updated: 2026-10-03