AI Gateway
Every AI request on one governed path
Warden is the gateway every AI request in your organization passes through, from the platform's own features and from outside tools such as Cursor, Claude Code and your own SDKs. One pipeline owns authentication, safety guards, routing, provider failover, budgets, caching and a per-call audit record, so a policy is enforced once instead of in every feature. You control AI spend, data exposure and provider risk from a single place.
When every application calls models directly, every application needs its own guardrails. Warden makes one governed path for all of them, from platform features to outside tools and your own SDKs, with routing, failover, budgets, guards, and an audit record for every call, including refusals, so a policy is set once and enforced everywhere.
- AI source pools with priority ordering and dynamic source selection
- Routing policies for cost controls, rate limits, and fallback strategies
- Content guards for PII handling and prompt filtering, applied by Shroud and Vigil in the request path
- Per-org, per-team, and per-user model preferences and overrides
- Routing log, cost analysis, and end-to-end AI usage analytics
What you get
What Warden gives your team.
Drop-in compatible APIs
OpenAI- and Anthropic-compatible APIs, streaming included. Existing SDKs, LangChain and IDE assistants point at Warden by changing a base URL and key.
Scoped Warden Keys
Assign keys to a user, team, service or organization, with IP allow-lists, scheduled rotation and per-key limits on pools, models and MCP servers.
Pools with automatic failover
Priority-ordered or load-balanced AI pools with health-aware failover and session-consistent routing across providers.
Budgets and rate limits
Hierarchical budgets warn, block, or reroute to a cheaper pool at a threshold. Rate limits and concurrency caps apply per organization, team, user, key, pool and source.
Response caching
Exact and semantic response caching through cache banks, alongside provider-side prompt caching and savings recommendations.
Audit for every call
Every call, including denials and provider errors, writes a record of the guards that ran, the source chosen and the cost, with spend and provider-health dashboards.
How it works
- 1
Group model sources into pools
Register your model providers as sources, then gather them into pools. Choose priority-ordered or load-balanced selection, with health-aware failover and sticky routing for consistent sessions.
- 2
Issue scoped Warden Keys
Create keys for a person, team, service or organization. Limit each to approved pools, sources, models and MCP servers, add IP allow-lists, and schedule rotation with a grace window.
- 3
Point existing tools at Warden
Change the base URL and key in an SDK, LangChain app or IDE assistant. Requests, streaming included, now pass through one pipeline of guards, routing, caching and budgets.
- 4
Watch spend and routing decisions
Set budgets that warn, block or reroute to a cheaper pool. Review the dashboard and decision log, where every call, including refusals and provider errors, has an audit row.
Ask AI about Warden
Every Warden page has an Ask AI button. It opens a chat that already has the page, its rows, and your selection. Try asking:
- Which Warden Keys on this page are assigned to services rather than to people?
- Which budgets listed here are closest to their threshold, and what happens when each is reached?
- Why was this call denied in the routing decisions shown, and which guards ran?
How teams use it
- Give every developer tool one governed endpoint and key instead of raw provider credentials.
- Cap a team's monthly AI spend and reroute traffic to a cheaper pool when it nears the limit.
- Keep work running through a provider outage with automatic failover.
Frequently asked questions
See Warden in action
Join the early-access list, or talk to our team about your environment.
Talk to us