One key and one model-agnostic endpoint for every agent you run — Claude Code, Cline, Aider, OpenCode, or your own. OpenAI-compatible, with native Anthropic Messages.
OpenAI · Anthropic · xAI · Moonshot • Try it for free
The problem
A person who sees "service unavailable" waits a minute, checks a status page, or fixes the typo. An agent can't. It obeys the status code, and a 503 says try again, so it retries into its budget.
A typical gateway, for a typo'd model id
HTTP/1.1 503 Service Unavailable
{
"detail": "service unavailable, contact support"
}Reads as an outage. The agent backs off and tries again, and again.
Gatewayz, for the same request
HTTP/1.1 400 Bad Request
{
"error": {
"type": "invalid_request_error",
"code": "model_not_found"
}
}A client error with a stable code. The agent stops, and your logs say why.
How we build it
Each one is something you can check against the API, not a promise on a page.
We don't sell a model of our own, so we have no reason to steer you toward one.
An alias resolves to its exact snapshot. An unknown id gets a 400, never a substitute you're billed for.
"claude-sonnet-4-5" -> claude-sonnet-4-5-20250929
"claude-sonet-9" -> 400 model_not_foundNobody is watching when an agent runs, so every failure has to say what to do next in the status code itself — stop, or try again.
Ask for a model that doesn't exist and the answer is 400 model_not_found — a typo in the request, not an outage. Spend past a key's cap and the answer is 402 request_cap_exhausted, and an empty balance gives 402 insufficient_credits. Neither clears by waiting, so an agent stops instead of retrying into its budget.
The two worth retrying look different on purpose: 429 when you are rate limited, and 5xx when a provider fails. And if a provider dies mid-answer, the stream ends with an explicit error event rather than simply stopping, so a half-finished reply is never mistaken for a complete one.
Stop, or retry
/health/quick names the build serving you.Under the hood
01
Agent
Your code, Claude Code, or any OpenAI-compatible client
02
Key and cap check
Valid key, credits left, under its spend cap
03
Exact model resolution
Alias to snapshot, or a 400
04
Provider
Same-model failover, never a different model
05
Metered response
Billed per token, served by a named build
Who it's for
Point Claude Code at native Anthropic Messages, or any agent at the OpenAI-compatible endpoint. Cap each key.
One balance across providers, and a key per workload with its own spend cap.
Software that spends on its own needs errors it can act on and a budget it can't exceed.
Run a community GPU node (beta).
Integrate
Two environment variables for Claude Code. One base URL for everything else.
export ANTHROPIC_BASE_URL=https://api.gatewayz.ai
export ANTHROPIC_AUTH_TOKEN=<your Gatewayz key>
claudeClaude Code (native Anthropic Messages)
Running an organization made of agents? FlashyOS's agent CLI is built on Gatewayz
Your own agent works the same way: set the base URL, keep your code.
from openai import OpenAI
client = OpenAI(
base_url="https://api.gatewayz.ai/v1",
api_key="<your Gatewayz key>",
)
reply = client.chat.completions.create(
model="anthropic/claude-sonnet-4-6",
messages=[{"role": "user", "content": "Hello"}],
)
print(reply.choices[0].message.content)Cline, Aider, OpenCode, Continue — via the OpenAI-compatible endpoint
Simple integration
Claude Code and every model in the catalog work with the Gatewayz API. Point any OpenAI-compatible client at one base URL and use one key. Change providers by changing the model string — the rest of your code stays as it is.
curl https://api.gatewayz.ai/v1/chat/completions \
-H "Authorization: Bearer $GATEWAYZ_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "anthropic/claude-sonnet-4-6",
"messages": [{"role": "user", "content": "Hello"}]
}'Prefer the Anthropic format? Native Messages lives at /v1/messages, with tool use and prompt caching passed through.
Autonomous organizations
An organization made of agents makes every decision through an inference call. Gatewayz is where those calls run.
FlashyOS's agent CLI ships Gatewayz as its only inference provider.
npx @flashyos/agent init --inference gatewayzAny autonomous organization on the FlashyOS network can run on the same endpoint.
Each agent or role gets its own key and request cap, so one runaway agent can't drain the rest.
An unknown model is a 400 and a spent cap is a 402. Neither looks like a failure worth retrying.
Tag a request with x-gatewayz-tag, then read calls, tokens and cost for that tag from /v1/usage.
Pricing
Security & data
Everything you need to know about the Gatewayz inference layer
Still have questions? Contact our team →
Create a key, set a cap, and point your agent at the gateway.