AI gateways
Bonito vs LiteLLM (BerriAI)
Self-run open-source proxy vs. a managed control plane
LiteLLM is an open-source, OpenAI-compatible gateway and SDK for 140+ providers, with a rich router, budgets, an MCP gateway, and an Enterprise tier with SSO, audit logs and SOC 2 Type 2. You run it yourself. Bonito is the managed option: the gateway plus agents, knowledge bases and residency pinning, without operating the proxy.
LiteLLM pricing
Open source: free, self-hosted. Enterprise: pricing not published, sold directly and through AWS and Azure Marketplace.
Model costs
No markup. Self-hosted with your own provider keys.
What LiteLLM does well
- The widest open-source provider coverage, with fast support for new models
- A rich router: cost, latency, usage and least-busy strategies, weighted failover, cooldowns and retry policies
- Budgets and spend tracking per key, user, team and org, in open source
- Images and /v1/videos (Sora, Veo) through the same proxy and keys, plus an MCP gateway
- Semantic caching on Redis, Qdrant or Valkey
Where Bonito goes further
- Managed: no proxy to deploy, patch, scale or secure
- Agents with a visual canvas, chat builder, approval queue, schedules and persistent memory; LiteLLM's Agent Platform is an early open-source release on Kubernetes
- Automatic Bedrock cross-region inference profiles; LiteLLM needs you to name the profile id
- Managed knowledge bases with a BYO bucket
- Data residency pinning to CA, US or EU as one setting that also constrains failover
- Agent trace tree with cost per request, agent, team and provider, without wiring a separate observability tool
Where LiteLLM is ahead of Bonito
- Open source and fully self-hostable
- Semantic caching; Bonito has none
- Provider breadth (140+) and router strategy options
Capability by capability
| Capability | Bonito | LiteLLM |
|---|---|---|
| Agents | ||
| Build and run agents | Yes | Partialopen-source Agent Platform (May 2026); no visual builder |
| Approval queue and scheduled runs | Yes | Nonot found in their docs (Sep 2026) |
| Dollar budgets per agent | Yes | Partialbudgets per key, user, team and org |
| MCP support | Yes | Yes |
| Routing | ||
| OpenAI-compatible gateway | Yes | Yes |
| Automatic failover across providers | Yes | Yes |
| Cheaper models for simple work (cost routing) | Yes | Yes |
| Bedrock, Vertex, Azure and direct providers in one place | Yes | Yes |
| Semantic caching | No | Yes |
| Governance | ||
| SSO, RBAC and audit log | Yes | YesSSO beyond 5 users and audit logs on Enterprise |
| Data residency pinning | Yes | Partialregion filtering in the router you configure |
| Observability and knowledge | ||
| Agent traces with cost per request and agent | Yes | Partialspend by key, user, team and tag; traces via Langfuse, LangSmith or Arize |
| Knowledge bases (RAG) | Yes | Nonot found in their docs (Sep 2026) |
| How you run it | ||
| Whole stack from one config file | Yes | Yesconfig.yaml for the proxy |
| Self-hosted or open source | Partial | Yes |
| No markup on inference | Yes | Yes |
Checked against LiteLLM's own documentation on 2026-09-27. Products change quickly; if something here is out of date, tell us at hello@trybonito.com and we will fix it.
Sources
See Bonito on your own stack
Start free, or get a report on where agents fit in your company first.