AI Gateway - Memory and model routing for any AI agent

Mastra's AI gateway handles memory, model selection, and failover automatically. Access 300+ models from every major provider with transparent pricing and no token markup.

Get Started for FreeDocumentation

Give any agent human-like memory

Compress chat history into compact memories that help agents remember what matters. Lower token usage and latency without losing important context.

Memory That Never Compacts

Stop interrupting users and losing important details with compaction. Form dense observations in the background without blocking the conversation or throwing anything away.

Observational Memory

Run two background agents — an Observer and a Reflector — that watch your agent's conversations and maintain a dense observation log. Replace raw message history as it grows.

5–40x Compression

Set when and how aggressively conversations compress. Configure observation and reflection thresholds to balance memory density against detail retention.

Prompt Caching

Append observations over time rather than rebuilding each turn. Keep the prompt prefix stable and cacheable — the longer the conversation, the more you save.

Route all traffic through one gateway

Use one endpoint for every model. Skip provider SDKs — one API key covers everything. Mastra's open-source AI gateway gives you a single integration point that routes to all major providers.

300+ Models

Access OpenAI, Anthropic, Google, Meta, Mistral, and hundreds more. Integrate once and reach every provider through a single gateway.

High Availability

Run on distributed infrastructure with automatic failover across providers.

Works with Any Stack

Change your base URL and start routing through the gateway. Use Python, TypeScript, any framework, any client.

Bring Your Own Keys

Plug in your own provider API keys for direct routing to OpenAI, Anthropic, Google, or any provider. Get the same memory features either way.

Transparent Pricing

Pay the same rate as the underlying provider with no markup on tokens. See exactly where your spend goes in the dashboard.

Start building with a free account, or explore our pricing for more

Full Pricing

Starter

Free for everyone

$0/ month
  • 100K observability events+ $10/100K
  • 24 CPU hours+ $0.35/hr
  • 15 days of data retention
  • Unlimited users, deployments, and projects
Get started for free

Teams

For growing teams

$250/ month
  • 1M observability events+ $8/100K
  • 250 CPU hours+ $0.25/hr
  • 6 months of data retention
  • Multiple teams, SSO, and SOC 2 docs
Get Teams

Enterprise

For teams at scale

Custom pricing

Custom volume and retention, with RBAC, audit logs, support and uptime SLAs, and a dedicated support engineer.

Contact sales

Mastra is powering the best AI teams

Case Studies

How Replit Agent 3 creates thousands of Mastra agents every day

Read case study

How SoftBank is restoring Japan's white-collar productivity using Mastra

Read case study

How Sanity Built a Content Agent That Actually Understands Your CMS

Read case study

How Marsh's globally distributed workforce uses LenAI agentic search, powered by Mastra

 

Frequently asked questions

What is AI Gateway?

Mastra's AI gateway adds observational memory and unified model routing to any agent. Use one endpoint for every model and one API key for every provider. Observational memory forms dense observations in the background without interrupting the conversation.

How does AI Gateway prevent agents from losing context?

AI Gateway prevents context loss by running two background agents that maintain a dense observation log without blocking the conversation or throwing anything away. An Observer watches conversations and compresses message history into concise notes about what happened. A Reflector condenses observations when they grow too long. Compression is typically 5-40x.

How does prompt caching work in AI Gateway?

AI Gateway appends observations over time rather than rebuilding the prompt each turn. Keeping the prompt prefix stable and cacheable means the longer the conversation runs, the more you save.

How does AI Gateway route traffic across model providers?

AI Gateway is the single integration point for every model provider your agents need to reach. One API key covers everything across more than 300 models including OpenAI, Anthropic, Google, Meta and Mistral, with no provider SDKs required. Change your base URL and start routing through the gateway using Python, TypeScript, any framework, any client. Plug in your own provider API keys for direct routing to any provider and get the same memory features either way. Run on distributed infrastructure with automatic failover across providers.

Does AI Gateway work with agents not built on Mastra?

AI Gateway works with any agent stack. Change your base URL and start routing through Mastra's open-source AI gateway using Python, TypeScript, any framework and any client. Observational memory and model routing apply to any agent regardless of how it was built.

Start building today

Route all your LLM traffic through one gateway built for production agents

Get StartedDocumentation