// the find
Klavis-AI/klavis
Klavis AI: MCP integration platforms that let AI agents use tools reliably at any scale
Klavis is an MCP integration platform with three products: prebuilt MCP servers for 100+ SaaS tools with OAuth flows, 'Strata' which multiplexes multiple MCP servers behind a single endpoint to reduce context bloat, and a sandbox environment for reproducible LLM training and RL. It's primarily a managed service with self-hosting available via their Docker images. Target audience is teams building AI agents that need to connect to real business tools without building and maintaining each integration themselves.
The OAuth handling is the real value here — managing OAuth flows for 40+ services (Salesforce, HubSpot, Google Workspace, etc.) is genuinely tedious work and they've absorbed it. The Strata multiplexer addresses a real problem: an agent with 500 tools loaded burns context budget before doing anything useful, and having a single MCP endpoint that surfaces only relevant tools is a practical solution. The sandbox dump/restore model is well-designed for RL training — you can initialize a Gmail or Jira sandbox to a known state, run an agent episode, then reset cleanly without hitting production systems. Multiple consumption modes (Docker, SDK, REST) means it fits into different architectures without forcing a rewrite.
The credential security story is essentially undocumented — when a user authorizes Gmail, it's not clear how tokens are stored, whether they're encrypted per-tenant, or what happens on a breach. For anything touching real email or financial accounts, that's not a detail you can skip. The self-hosting story is weaker than advertised: you're pulling their pre-built Docker images, not running a stack you can actually audit or modify. The Strata context optimization is described vaguely; there's no explanation of whether tool selection is semantic/dynamic or just static grouping, which matters a lot for how well it actually works. The README has a variable naming bug in the Python quickstart (`klavis_client` used after assigning to `klavis`), which is a small thing but suggests the docs aren't being run as real code.