Fermion is the runtime for agentic AI. Deploy tools, models, and long-running workflows behind one fast API — with sub-second cold starts, durable state, and observability built in.
1import { Fermion } from "@fermion/sdk";
2
3const agent = new Fermion.Agent({
4 model: "orion-4",
5 tools: [search, postgres, browser],
6 memory: "durable",
7});
8
9const run = await agent.stream({
10 input: "Draft the Q3 board update",
11});
Define once. Declare your agent — model, tools, and durable memory — in a few typed lines.
Run anywhere. Stream long-running workflows behind one fast API, with sub-second cold starts.
Observe everything. Trace every tool call, token, and dollar in real time — then replay any run.
Trusted by engineering teams shipping AI to millions
Stop stitching together queues, sandboxes, and dashboards. Fermion is the full runtime — from the first prototype to millions of runs a day.
Snapshot-restored sandboxes resume in 380ms, so idle agents still feel instant. Never pay for warm pools you don't need.
Route across frontier and open models with automatic failover, streaming, and per-token cost caps — no vendor rewrites.
Long-running workflows survive restarts, retries, and deploys. State is checkpointed, so nothing is ever lost mid-run.
Trace every tool call, token, and dollar in real time. Replay any run step by step and share a permalink with your team.
First-class TypeScript and Python with generated types for every tool. Go from a blank file to a deployed agent in an afternoon.
SOC 2 Type II, SSO, and private VPC networking. Your prompts and data are isolated per tenant and never train a model.
Every run is a distributed trace. Drill into spans, inspect the inputs and outputs of each tool call, and catch regressions before your users ever notice them.
Start free and scale when you do. No seat taxes, no annual lock-in, no surprises on your bill.
Free to start, sub-second cold starts, and full observability from your very first run.