On this page
The LiteLLM alternative
for agents that touch real systems.
Verosek is an AI and MCP gateway: keep your SDK, route model calls with fallback, set budgets and rate limits per key, connect your tools, and have each tool call decided by the key's rules before it runs. Moving from LiteLLM takes two values, base URL and key.

A full gateway on one key
What you expect from a gateway: routing, budgets, connectors and traces.
Route with fallback.
Verosek routes to OpenAI, Anthropic, Google, Azure, Bedrock, Ollama and any OpenAI-compatible server. Calls fall back in priority order.
Budgets and limits on the key.
Each key carries its allowed models, budget, rate limits, expiry and per-tool rules.
Connect your tools.
Connectors for databases, code hosts, ticketing, chat, payments and cloud, behind one MCP endpoint.
Run agents on it too.
Versioned agents, with their own key and limits.
Scanning comes with it.
Shield's checks ship in the gateway, with no second vendor to buy.
Why Verosek over LiteLLM?
Verosek is for platform teams whose own agents write to production systems. Audit logs, SSO past five users and guardrails per key or team need LiteLLM's Enterprise licence, priced through sales, and its secret detection needs the licence too, and scans the request only. All of that is on Verosek's Team plan, at a published $999 a month.
It reads the statement, not a pattern on its values.
Verosek reads the SQL statement itself, where LiteLLM's rules go by tool, parameter name and patterns on values. Verosek parses seven SQL dialects, so you can give an agent a production system and say exactly what it may do to it. LiteLLM's MCP access falls back to a wider list when the permission lookup cannot reach the database, as of version 1.104.0.
A record your security review can read.
Verosek seals one record for each request, model calls and tool calls together, with the reason for each decision, where LiteLLM's audit logs list key, team, user and model changes, on Enterprise, and request and response logs go out through callbacks. Give a regulator your public keys, and they can check each decision themselves.
Refuses by default. You choose the exceptions.
A blocking rule that cannot run refuses, and so does a request the record cannot hold. You can set one rule on a key to let calls through when the check itself breaks. LiteLLM's generic guardrail fails closed by default, and its built-in detection callback fails open on a non-HTTP error. Its MCP sub-calls do not inherit a request's guardrails unless one is on by default. With LiteLLM's fail-closed flags off, an outage of its counter store degrades budgets and rate limits to per-instance counters.
Attacks become findings, and findings become rules on the key.
Assay, an open MCP security benchmark we maintain, was presented at Black Hat India Arsenal. Our red-team framework holds 205 attack techniques, 68 for MCP, with findings anyone can recompute.
Does my proxy already do this?
Part of it, perhaps. Allow lists per key and deny by default are common. On one key, Verosek also routes your model calls, sets budgets, scans with no second vendor, reads the SQL statement down to its tables and row bounds, and seals the reason for each decision. LiteLLM's open-source gateway gives one endpoint for your MCP tools, with access by key and team. You can keep your proxy behind Verosek, which routes to any OpenAI-compatible server.
Start on the gateway you will keep.
There is no reason to stand up one gateway now and move your keys, budgets and rules later. The record of calls made before Verosek was in the path cannot be written afterwards, and a call that already ran cannot be taken back. Watch mode shows what a rule would have blocked before it enforces, so nothing stops until you say.
At a glance
| The point | Verosek | LiteLLM |
|---|---|---|
| Model routing | OpenAI, Anthropic, Google, Azure, Bedrock, Ollama and any OpenAI-compatible server, with fallback | One API to over 140 providers, with fallbacks in open source |
| Budgets and limits | Allowed models, budget, rate limits and expiry on each key | Budgets and rate limits in open source |
| Tools and connectors | A catalogue of connectors, with OAuth where the server needs it | One endpoint for your MCP tools, in open source |
| Tool rules | Per key and per tool, before each call runs: a SQL statement's operation, tables and row bounds, a payment's amount, a file path, a namespace, a recipient | Allow lists at six levels, per key and team among them, with patterns on values |
| What makes a block | A named rule. By default, a model's score alone only flags. | Its permission rules and guardrails; its docs say guardrails are classifiers, so a crafted injection can pass any of them |
| Scanning | Prompt injection, personal data and secrets in prompts, replies, tool arguments and tool results, on both plans, with no second vendor to buy | A guardrail framework in open source, with over fifty integrations, from its own Content Filter to hosted detectors; guardrails per key or team on Enterprise |
| The record | One sealed record for each request, with the reason for each decision | Request logs through callbacks, in open source; admin audit logs on Enterprise |
| Human approval on a tool call | A named person decides the calls you choose | No server-side approval step per tool call, as of version 1.104.0 |
| On a published plan | Team at a published $999 a month, with every Shield profile, sealed records and SSO | Routing, budgets and guardrails in open source; SSO past five users, audit logs and guardrails per key or team through sales |
| Where it runs | Enterprise: private cloud, VPC or air-gapped, deployed with us | Open source: self-hosted, with an MIT core. Enterprise: air-gapped too, or hosted through sales |
As of 5 Nov 2026.
Is Verosek a drop-in replacement for LiteLLM?
An app on an SDK.
For an app that calls a gateway through an OpenAI, Anthropic or Gemini SDK, yes: you change the base URL and the key, and your calls keep their shape.
Code on the Python library.
Code that imports LiteLLM's Python library moves to an OpenAI, Anthropic or Gemini SDK first.
Virtual keys and budgets.
Each virtual key becomes a Verosek key, with its budget and limits, and we set that up with you. Each team or agent gets its own key and its own budget.
Rules per team or per user.
Verosek's rules live on the key, and a key can bind tool calls to the end user it names, so rules set per team or per user move onto the keys they hold.
Detectors you added.
LiteLLM ships over fifty guardrail integrations, from its own Content Filter to hosted detectors. Shield's checks ship in the gateway, so a detector you pay for separately can go.
Logs through callbacks.
LiteLLM sends request and response logs out through callbacks. Every request becomes a trace of typed steps, scrubbed of secrets and personal data, then sealed.
Spend and traces.
The console shows each key's spend and each request's trace.
When LiteLLM is enough
If your agents only call models and never change anything, a router may be enough.
Frequently Asked Questions
Three things: what sits behind an Enterprise licence, what happens when a check errors, and whether anyone outside can check the record. Verosek answers all three on the Team plan, at a published $999 a month. It can also hold the calls you choose for a named person.
Verosek routes to OpenAI, Anthropic, Google, Azure, Bedrock and Ollama by name, and to any other provider through an OpenAI-compatible server.
Verosek is closed source. The gateway and its admin console hold no licence check, telemetry, analytics or call to a Verosek host.
Team is $999 a month, billed yearly. Enterprise has custom terms. Both plans start with a call. See the plans. LiteLLM's audit logs need an Enterprise licence, priced through sales. As of 5 Nov 2026.
Verosek is the gateway as well: one key routes your model calls, Bedrock among them, sets budgets and scans with no second vendor, and it reads the SQL statement and seals the reason for each decision. AgentCore Policy goes by user identity, a tool call's input parameters, earlier steps in the session and signals such as a guardrail's score; its docs do not say a policy reads a SQL statement. As of 5 Nov 2026.
An app leaves by changing the same two values. On a self-hosted deployment, on the Enterprise plan, the gateway, its signing key and your records stay with you.
Bring one real agent to a call, and we stand up a gateway for it.
Sources
- AWS docs: policy https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/policy.html Read 6 Nov 2026
- LiteLLM https://www.litellm.ai/ Read 5 Nov 2026
- LiteLLM docs: audit logs https://docs.litellm.ai/docs/proxy/audit_logs Read 5 Nov 2026
- LiteLLM docs: enterprise https://docs.litellm.ai/docs/enterprise Read 5 Nov 2026
- LiteLLM docs: generic guardrail api https://docs.litellm.ai/docs/adding_provider/generic_guardrail_api Read 5 Nov 2026
- LiteLLM docs: llms.txt https://docs.litellm.ai/llms.txt Read 5 Nov 2026
- LiteLLM docs: mcp https://docs.litellm.ai/docs/mcp Read 6 Nov 2026
- LiteLLM docs: mcp control https://docs.litellm.ai/docs/mcp_control Read 5 Nov 2026
- LiteLLM docs: mcp guardrail https://docs.litellm.ai/docs/mcp_guardrail Read 5 Nov 2026
- LiteLLM docs: security owasp llm top10 https://docs.litellm.ai/docs/proxy/security_owasp_llm_top10 Read 5 Nov 2026
- LiteLLM docs: tool permission https://docs.litellm.ai/docs/proxy/guardrails/tool_permission Read 5 Nov 2026
- LiteLLM on GitHub: LICENSE https://github.com/BerriAI/litellm/blob/main/LICENSE Read 5 Nov 2026
- LiteLLM: pricing https://www.litellm.ai/pricing Read 5 Nov 2026

