AI governance
AI governance and policy enforcement for LLMs and agents
Policies in a document don't stop a request. Leanroute applies your rules on every model and tool call, in one place, before anything reaches a provider: what data may be sent, which tools agents may use, how much can be spent, and a record of what happened.
Guardrails on what goes in and what comes out
A seven-rule library: PII redaction, secret detection, prompt injection, moderation, data exfiltration, profanity and language. Add your own patterns. Each rule blocks, redacts or only logs, and rules can check responses as well as prompts.
Guardrail library →Control which tools agents can reach
MCP servers must be on your organization's allowlist before the gateway will connect to them. A request naming any other server is refused. Agents can list the servers they are allowed to use.
MCP allowlist →Spend and rate limits as policy
Hard daily and monthly caps and requests- and tokens-per-minute limits, for the whole organization and for each key, enforced before any provider is called.
Cost control →Separated roles and keys
Runtime keys send model traffic but can't manage the account. Admin keys manage keys and read usage but can't send model traffic. Top-ups and provider keys can only be changed in the dashboard. Team members are owners, admins or members.
Security →Data handling you can explain
The request log stores who, what model, tokens and cost, never prompt or response content. A no-persistence mode turns off caching for your organization. Provider keys are encrypted (AES-256-GCM). Hosted in Singapore; DPA on request.
Data handling →A record of every request
Every request is logged with key, model, provider, cost and status, and agent traffic with its run, agent name and the tools it called. Read it in the dashboard or export it through the API.
Agent runs →Honest limits
What a governance review should know before choosing us:
- No SOC 2 report yet. We run automated security scans monthly; details on the security page.
- No SSO or SCIM yet, and no export of an audit trail of admin actions.
- Policy is per organization today. Per-key tool allowlists and policy over an agent's own function tools are on the roadmap.
- Guardrails check prompts and responses. They don't yet inspect MCP tool arguments or tool results.
- Guardrails are pattern and moderation based. They reduce risk; they are not a complete data-loss-prevention system.
Reviewing agents for production? Use the AI agent governance checklist: 50 controls across inventory, access, data, tools, spend, audit and incidents, with which ones a gateway can enforce.
Questions
- What is AI governance at the gateway level?
- Applying the same rules to every model call in one place, instead of in each application: what data may be sent, which models and tools may be used, how much may be spent, and a record of what happened. A gateway sits on every request, so it can enforce those rules before a provider is called.
- Can I block PII from reaching an LLM provider?
- Yes. The PII redaction rule masks emails, phone numbers and similar data in prompts before they leave the gateway, or you can set it to block the request instead.
- Does Leanroute store our prompts?
- Not in the database. The request log records metadata only. Prompts are held briefly in the cache to serve repeat requests, and the no-persistence setting turns that off.
- Is Leanroute SOC 2 certified?
- Not yet. See the security page for what is in place today, and ask us for the DPA and data-handling details your review needs.
Turn on guardrails and caps in minutes. New accounts get 100 free requests, no card needed.