Serving, the way agents need it.
Most inference is served for chat: one question, one answer, goodbye. Layer X1's engine was built from the ground up for agents — runs that last hours, hammer tools, and cannot afford a dropped call. You point an agent at the endpoint; the engine guarantees the rest.
Five guarantees, held on every request.
You point an agent at an endpoint. We do the rest. You never see the machinery — you see what it guarantees.
- 01
Always on
Degradation never reaches your run — serving recovers before your agent notices, with no retries to write and no fallback lists to maintain on your side. Zero downtime is the design target, not a status-page apology.
- 02
Fast for agents, from the ground up
Chat inference optimizes one answer. Agent inference is different: hours-long trajectories, tool calls every few seconds, context that keeps growing. Our stack was built for that shape of traffic from day one — sustained throughput across the whole run, not a fast first token and a slow everything else.
- 03
Usage that reads huge
Plans carry a monthly pool of included usage measured in dollars at public list rates — $60 of model usage on a $5 plan, $300 on Pro. Nearly every model in the catalog draws from that one pool on every paid plan: no per-model surcharges, nothing to tune. A couple of frontier models are tier-reserved (Opus 5 on Pro/Max, Fable 5 on Max) and marked on the catalog with a plan chip.
- 04
Tool calls that hold
Agent loops live and die on tool calling. The engine enforces strict, schema-faithful tool calls under load, on every model in the catalog, in both agent dialects — so a malformed call never silently derails a run.
- 05
Any agent, no ceremony
Anthropic-compatible and OpenAI-compatible endpoints, streaming everywhere, one key. If it runs an agent loop, it runs here — the engine does not care which framework is driving.
One door. Four protocols.
Whatever dialect your agent already speaks, the endpoint answers it — same key, same catalog, same guarantees. Pick the path your tooling expects and point it here.
Messages
POST /v1/messagesThe dialect agent harnesses like Claude Code speak natively. Swap one base URL and the loop lands here — streaming included.
API reference →OpenAI-compatibleChat Completions
POST /v1/chat/completionsThe dialect nearly every tool ships with. Same key, same models, no SDK changes — if it talks to a chat endpoint, it runs here.
API reference →OpenAI-compatibleResponses
POST /v1/responsesThe newer agent-native dialect used by harnesses like Codex — structured turns and tool calls, streamed end to end.
API reference →OpenAI-compatibleEmbeddings
POST /v1/embeddingsVectors for the retrieval side of your stack, served behind the same key and the same door as everything else.
API reference →Every model, behind one endpoint.
You hold one key and one endpoint — never another vendor account, quota, or bill. Open-weight flagships sit next to premium frontier models behind the same door, and the catalog keeps growing. One key reaches all of it.
from openai import OpenAI
client = OpenAI(
base_url="https://api.layerx1.com/v1",
api_key="lx1_your_key",
)
response = client.chat.completions.create(
model="lx1-gpt-oss-120b",
messages=[{"role": "user", "content": "Say hello."}],
)lx1_ key · every model in the catalogFull API reference →Point any agent here in one command.
Run npx layerx1 and pick your tools — the installer writes each config where the tool expects it, key and model included; GUI-configured tools get their exact paste-in values. No SDK, no code changes. The session below is the setup.
Your agent stays the same.
Its inference gets better.
Start free · no card · Starter from $5/mo