"We replaced three vendor SDKs with one endpoint. The part I did not expect: the runtime says no fast. Bad requests die in 300ms instead of burning a minute of tokens."
Maya R. — Product Lead
The operations layer for every model
Deterministic by design. Every routing decision is a rule, not a guess.
- 100M+
- daily tokens
- 450+
- models
- 100+
- providers
- ~300ms
- decision p50
how it works
One instruction in. A clean, structured result out.
Whatever you send — a single prompt, a multi-agent task, or a scheduled workflow — the gateway decides how it should run, and hands back an answer, not a routing problem.
alphaproxy — route
$alphaproxy route --optimize quality,cost --prompt "compare five competitors and find the sharpest positioning gap"
decisiontyped · task=analysis · tier 2 · 118ms · 0 tokens
candidates evaluated14 models · 9 providers
selectedclaude-sonnet-4.6 · anthropic · fastest healthy host
fallback chaingpt-5.1 → gemini-3-pro · health-routed
cost vs. direct−68%
latency1.8s to first token
Routing is a typed decision, not a guess: the decision layer classifies the task, then picks the model, host and fallbacks per request — your integration never changes.
alphaproxy — exec
$alphaproxy exec --workflow "audit sales process, map friction points" --agents 3
plan3 ephemeral agents · parallel · isolated sandbox
research-agentpulled CRM + call notes · mid tier
analysis-agentscored 6 handoff stages · efficient tier
report-agentdrafted findings doc · frontier
shared context3 agents · 1 session · nothing re-explained
completion gatepassed · doc opens, figures reconcile
artifacts produced1 report · 1 chart
Agents are spun up per task and torn down after. Each gets the right-sized model, tools and skills for its piece of work, and the completion gate checks the merged result before you see it.
alphaproxy — schedule
$alphaproxy schedule --trigger weekly --workflow "founder update from product, revenue, hiring notes"
sources connecteddocs · sheets · linear
next runMon 09:00 · weekly · survives restarts
model policyquality-first · routed per run
guardrailscontent gate · budget cap · retry policy
deliveryemail + Slack
traceper-run log · cost · model
statusarmed
Describe the outcome once. It becomes a durable job that runs through the same gateway, key and cost controls as any live request — and reports what it did and what it cost.
under the hood
Decisions are typed. Generation only happens when generation is the answer.
Every request hits the decision layer first. Most of what a request needs — is this safe, which lane, does it need a sandbox — is answered as a typed value, not a paragraph of reasoning. Generation is reserved for the part of the task that actually needs it, and even then it's split across the right-sized model for each piece of work.
decision layer / final response
model assigned per subagent
request flow
- Decision layer · decides
- Classifier-driven decisions from a non-generative evaluator. Choice, score and open questions come back as typed values — no generated text, ~300ms p50, a fraction of a cent per million tokens. Nine decision seams in production.
- Verifiable-reward generation · generates
- When output must be produced, the right-sized model produces it — and the same decision layer verifies the result against checks that can actually pass or fail: code that runs, files that open, deliverables that match the ask.
the gateway
Build once. Switch models anytime.
Every model, agent, and workflow behind a single endpoint. Swap providers, blend models, or route around an outage — without touching a line of your application.
topology
- your app
- your agents
- your workflows
ALPHAPROXY
1 key · 1 endpoint
decide · route · verify
- openai · anthropic
- google · mistral
- meta · deepseek
- + 100 more providers
−
base_url = "https://api.openai.com/v1"+
base_url = "https://api.alphaproxy.ai/v1"model = "auto" // or any of 450+- Unified access. One OpenAI-compatible endpoint reaches every leading model — nothing new to learn, one key to manage.
- Automatic routing. Each request is matched to the model best suited for quality, latency or cost — per call, not per integration.
- No vendor lock-in. Move off any single provider instantly; your code, prompts and key stay exactly the same.
- Agent-ready. Multi-step and multi-agent tasks keep context across the whole chain — sandboxes, browsers, files and 1000+ connectors included.
run an intent
Say what you want. Watch the state change.
Pick a line. The runtime shows how it would decide, route and deliver — before a single token is generated.
alphaproxy — tracestate
system log
Sessions that changed state.
"I asked for a market map and got a sourced deck, a spreadsheet and the raw research. Everything it produced I could verify. That is the whole point."
Daniel K. — Investor
"It read our repo, reproduced the regression in a sandbox and shipped a diff with tests. The completion gate refused to call it done until CI was green."
Priya S. — Engineering Manager
"Model choice stopped being a meeting. We say 'auto', the gateway picks, and our cost per task dropped by two thirds."
Elena V. — Strategy Director
"The Monday report writes itself now — pulled from Notion and Sheets, built into a deck, posted to Slack. Every run leaves a log I can audit."
Omar A. — Operations Lead
"One key, every model, agents that actually finish. We shipped our MVP through it and never wrote a provider integration."
Lucas P. — Founder
state transition
This is not a website.
It is a state.
Open a session. Get a key. Point your stack at one endpoint and let the runtime decide what runs.
free to start · no card · openai-compatible