We build the control layer for AI in software.
Worksi AI Control is one place to see, govern and account for every AI agent, chat assistant and copilot a company runs, without rewriting the applications that use them.
AI shipped faster than the tools to run it.
Teams ship agents and AI features every week. What usually follows is familiar.
- Nobody has the full inventory.
- Which agents, models and tools run in production, who owns them, and what appeared outside the approved list.
- Failures show up as complaints.
- A failing tool, a looping agent, a provider outage or a cost spike reaches customers or the invoice before it reaches an engineer, without evidence of what went wrong.
- Recovery is slow and risky.
- Pausing an agent or stopping a run should not need a code change and a deploy.
- Governance is an afterthought.
- Model allowlists, budgets, approvals for risky actions and an audit trail are hard to add later.
- Nobody can say what it is worth.
- Spend is not connected to outcomes, so value is argued rather than shown.
Principles you can see in the product.
Each one is a design decision you run into when you use the product.
- Work with what teams already run
- An OpenAI‑compatible gateway that also speaks Anthropic’s Messages API, OpenTelemetry ingestion and MCP. Connecting is a base URL change, not a rewrite.
- Your model account, your bill
- Customers bring their own provider keys. We never resell or mark up model tokens.
- Evidence before opinions
- Diagnosis runs deterministic checks first. AI reasoning is optional, scored for confidence and must cite evidence.
- Observe first, then enforce
- Every policy can run in monitor mode and be simulated against real traffic before it blocks anything.
- People decide the risky things
- High‑risk recovery actions and sensitive tool calls wait for a person, with four‑eyes when it matters.
- Honest numbers
- We publish our own measurements with the conditions behind them, and no customer logos or quotes we don’t have.
Built for the hot path.
Latency-sensitive traffic runs through a small Go data plane. Analysis and the control plane run beside it, never in the way of a request.
- Gateway and ingestion
- Go, streaming-safe, with telemetry that never blocks a request
- Control plane
- Python (FastAPI), deterministic detectors, LLM‑assisted diagnosis
- Telemetry storage
- ClickHouse
- Operational data
- PostgreSQL with pgvector for similar-incident search
- Messaging and counters
- Redis streams, pub/sub and atomic counters
- Console
- Next.js and TypeScript, light and dark themes
- Hosting
- Amazon Web Services
Company
- Company
- Worksi
- Website
- worksi.ai
- info@xeoo.io
See it on your own traffic.
Start a 3‑day trial with Pro features and no card, create an API key, and ask the gateway which models your key can use.
Terminal
curl -s https://gateway.worksi.ai/v1/models \ -H "Authorization: Bearer $AIW_API_KEY"