Klavis AI

Executive Summary

Klavis AI is an open source and hosted MCP integration platform: 300+ services exposed to AI agents as managed MCP servers with OAuth handled by the platform, so agents get authenticated access to Gmail, Salesforce, YouTube, and the rest of the SaaS long tail without custom auth plumbing. ^[1] Its signature product is Strata, a single unified MCP server that solves tool overload by progressively disclosing tools — agents discover services, then categories, then action names, and only load a full parameter schema at execution time — which Klavis reports beats the official GitHub and Notion MCP servers by +15.2% and +13.4% pass@1 on the MCPMark benchmark. ^[2]

The company is young and small: founded 2025 by Xiangkai Zeng (ex-Google DeepMind, Gemini) and Zihao Lin (ex-Lyft) out of Y Combinator's Spring 2025 (X25) batch, with a team of three and seed backing from HSG (formerly Sequoia China) and YC. ^{[3] [1]} The Apache 2.0 repo has 5.7K+ stars as of June 2026, and the platform claims SOC 2 Type II and GDPR compliance plus testimonials from the MCP project, CrewAI, LlamaIndex, Together AI, and Fireworks AI. ^{[4] [1]} Its homepage now leads with a second act: live, ultra-realistic sandbox environments — browser, code repos, SaaS tools, MCP servers — for training and evaluating agents. ^[1]

---

Product Overview

Klavis sits between agents and third-party SaaS APIs. Developers (or end users, via hosted auth flows) connect accounts once; Klavis manages the OAuth dance and token lifecycle and exposes each service as a production-ready MCP server — hosted or self-hosted — that any MCP client or agent framework can call. ^[1] CrewAI's testimonial highlights the core value proposition: integrations without manually coding OAuth flows. ^[1]

Strata, launched September 2025, is the layer that makes large catalogs usable. Instead of dumping hundreds of tool schemas into the prompt, Strata guides the agent through a staged discovery flow and can also wrap any external MCP server — custom, official, or community — converting it into the same discovery-driven architecture with no extra configuration. ^[2] Klavis reports 83%+ task accuracy on complex multi-app workflows with this approach. ^[2]

The newest surface is agent training: live sandbox environments with managed auth, seeded state data, state export, and parallel isolation for training and evaluation runs — aimed at teams doing RL and long-horizon agentic task evaluation across browsers, code repositories, SaaS tools, and MCP servers. ^[1]

Key Capabilities

Product Surfaces

---

Technical Architecture

Klavis is a primarily Python codebase (Klavis-AI/klavis, Apache 2.0, created April 2025) containing the MCP server implementations that also power the hosted platform. ^[4] Agents connect over standard MCP, so the platform is client-agnostic — anything that speaks MCP (Claude, Cursor, agent frameworks) can use it.

Strata's progressive discovery exposes four meta-tools instead of N tool schemas: discover_server_categories (which services are relevant), get_category_actions (action names and descriptions only), get_action_details (full schema for the chosen action), and execute_action (run it) — with optimizations that skip stages for simple integrations. The full parameter schema enters the context window only at execution time, which is how the catalog scales without overwhelming the model; on MCPMark this yields +15.2% pass@1 over the official GitHub MCP server and +13.4% over the official Notion server. ^[2]

Key Technical Details

---

Strengths

• Context-window economics as the differentiator — Strata's staged discovery loads schemas only at execution time, with benchmarked accuracy gains (+15.2% / +13.4% pass@1 on MCPMark vs official GitHub/Notion servers) rather than marketing claims ^[2]
• Genuinely open — Apache 2.0 (not a source-available license), 5.7K+ stars, 545 forks; you can self-host the servers and Strata outright ^[4]
• MCP-native from day one — built on the protocol rather than retrofitting a proprietary tool API, and Strata can wrap any third-party MCP server into its discovery flow ^[2]
• Credible ecosystem validation — testimonials from the MCP project itself plus CrewAI, LlamaIndex, Together AI, and Fireworks AI; SOC 2 Type II for enterprise procurement ^[1]
• A second product with RL tailwinds — live sandbox environments for agent training/eval ride the same infrastructure and target the growing RL-environments market ^[1]

---

Cautions

• Very early company — founded 2025, three-person team per YC; key-person and execution risk is real ^[3]
• Undisclosed seed — HSG and YC backing is credible, but no public round size means runway is unknowable from the outside ^[1]
• Smaller catalog than the leader — 300+ services versus Composio's 1000+ toolkits; long-tail coverage may require contributing your own servers ^[1]
• Pricing opacity — a free tier and paid/enterprise plans exist, but limits and meters are not comprehensively published, complicating cost forecasting ^[5]
• Focus split — the homepage now leads with training sandboxes rather than integrations; a three-person team running two product lines (integration platform + RL environments) is a bet on focus ^{[1] [3]}
• Open-issue backlog — 285 open issues on the main repo as of June 2026 suggests maintenance load is outpacing the small team ^[4]

---

Pricing & Licensing

Licensing model: Apache 2.0 — fully permissive; self-host, modify, and even commercialize without copyleft or service restrictions. ^[4]

Hidden costs: Self-hosting shifts OAuth app registration, token storage, and server operations onto your team; hosted-tier limits are not publicly itemized, so verify quotas before committing. ^[5]

---

Competitive Positioning

Direct Competitors

When to Choose Klavis AI Over Alternatives

• Choose Klavis when: agents need many tools and context-window bloat or tool-selection accuracy is the binding constraint — Strata is the most developed answer to that specific problem
• Choose Klavis when: a permissive Apache 2.0 license matters (Composio's platform is more commercial; Nango is ELv2 source-available)
• Choose Composio when: you want the largest ready-made toolkit catalog with sandboxed execution and a heavily funded vendor
• Choose Nango when: you also ship customer-facing product integrations (syncs, webhooks) and want one layer for both

---

Ideal Customer Profile

Best fit:

• Agent builders on MCP-native stacks (Claude, Cursor, CrewAI, LlamaIndex) who need authenticated access to dozens of SaaS tools without OAuth plumbing
• Teams hitting tool-overload limits — degraded accuracy or blown context budgets from large tool catalogs
• Companies requiring self-hosting under a permissive license, or SOC 2-vetted hosted infrastructure
• Labs and agent teams that also need realistic, stateful environments for RL training and evaluation

Poor fit:

• Teams that need the broadest possible pre-built catalog today (Composio's 1000+ toolkits)
• B2B SaaS companies whose primary need is customer-facing product integrations rather than agent tool access
• Buyers who need published, predictable pricing before procurement

---

Viability Assessment

Klavis has out-engineered its weight class — Strata's progressive discovery is the clearest articulated answer to MCP tool overload, with benchmark numbers to back it. The risk is structural, not technical: a tiny seed-stage team running two product lines against much better-capitalized competitors, in a category where the protocol itself (and first-party vendors) keeps absorbing functionality.

---

Bottom Line

Klavis AI is the open source, MCP-native bet in the MCP integration platform category: 300+ managed-OAuth integrations under Apache 2.0, and in Strata, the best-evidenced solution to the tool-overload problem — progressive discovery that benchmarks above the official GitHub and Notion MCP servers. ^{[1] [2]}

Recommended for: MCP-native agent teams constrained by context windows or tool-selection accuracy, and anyone who wants Composio-style managed tool access with a permissive open source license.

Not recommended for: Teams that need the largest catalog or a deep-pocketed vendor today (see Composio), or product-integration use cases (see Nango).

Outlook: Promising but fragile — the technical differentiation is real and the ecosystem endorsements are credible, but seed-stage scale and undisclosed funding make vendor durability the main question.

---

Research by Ry Walker Research • methodology