Archestra.AI Raises $10M Seed to Guard AI Agents’ Data Access

Archestra.AI, a London-based startup that lets companies plug AI agents into sensitive data without losing control, has raised a $10 million seed round led by 20VC, the fund run by podcaster and investor Harry Stebbings. The deal takes the firm’s total funding to $13.5 million and puts four Fortune 500 clients into production with a team of just seven people.

By the standards of its category, $10 million is a small cheque. It still points to something bigger across enterprise IT: large companies want agent security bought from a vendor other than the firm that built their AI model.

The $10M Round and Who Backed It

The seed money came in behind a roster of investors who know enterprise software from the inside. That detail matters more than the headline figure, because Archestra is selling into buyers who trust references over pitch decks.

The Investor List

20VC led the round as a new backer. Alongside it sat 20 Product, Visible Ventures, Tenacity Capital and Commit Fund. The angel list reads like a who’s-who of people who have actually shipped data tooling at scale:

• Olivier Pomel, chief executive and co-founder of the observability company Datadog
• Kieran Flanagan, a senior executive at the marketing-software firm HubSpot
• Carolyn Everson, a board member at private-equity firm Permira and a former advertising chief at Meta

The round follows a $3.3 million pre-seed in August 2025 led by Concept Ventures, which brought the lifetime total to $13.5 million. The new cash goes into engineering and go-to-market hiring, faster enterprise deployments in regulated industries, and the company’s open-source community.

The Founders’ Track Record

Archestra was founded in 2025 by Matvey Kukuy and Ildar Iskhakov, both alumni of Grafana Labs, the open-source analytics and visualisation company. The pair are repeat operators. They previously built Amixr, which Grafana Labs acquired in 2021, and Kukuy went on to co-found the alerting startup Keep, picked up by the search company Elastic in May 2025. Selling two infrastructure companies before this one is part of why a seven-person team can land Fortune 500 logos at all.

How Archestra Brokers Access to Sensitive Data

Most enterprises today run AI agents on a short leash. They let the agents draft emails or summarise tickets, and keep them well away from legal files, HR records or customer data, because an agent that hallucinates or gets hijacked could leak the lot. Archestra sells itself as the thing that sits in the middle and makes the leash unnecessary.

What an MCP Orchestrator Does

The platform is built around the Model Context Protocol (MCP), the open standard, backed by Anthropic, OpenAI, Google and Microsoft, that lets AI agents plug into external tools and data sources. You can read the open Model Context Protocol specification for the wiring details. Archestra acts as a broker over that protocol. Before an agent reaches a file, the platform checks identity and policy, applies guardrails against prompt injection and data exfiltration, and decides which dataset the agent is actually allowed to see. It ships as open source and connects to the tools enterprises already live in, from Jira and Confluence to SharePoint, Google Drive and Salesforce, and works with Claude, ChatGPT, Gemini and open-weight models alike. It belongs to the same wave of money flowing into agent-native infrastructure for the next generation of AI systems.

The Numbers Behind the Pitch

Kukuy frames the product as removing a forced choice. “Until now, AI teams had to choose between harshly restricting AI agents and building truly agentic systems, thereby exposing the business to enormous risks,” he said. “Archestra.AI eliminates that compromise.” The traction figures the company points to:

• 900+ MCP servers evaluated and supported
• 45 milliseconds of added latency at the 95th percentile
• Up to 96% lower inference cost through routing and in-process retrieval
• 3,700+ GitHub stars and 57 outside contributors on the open-source project

Why Buyers Want a Control Layer Model Makers Don’t Own

Here is the part of the story that outlives this one round. Kukuy says the Fortune 500 buyers he talks to do not want to hand the job of policing their agents to the same companies that sell them the models. The worry is vendor lock-in and a conflict of interest baked into the relationship.

Enterprises are looking for solutions, especially solutions engaging their data and managing their agents, which will be independent.

That was Kukuy, speaking about why large companies are reluctant to rely on model providers like Anthropic and OpenAI to keep agents safe across every business unit. His pitch for what an independent layer unlocks is blunt: “Archestra.AI lets agents do more. It lets agents work with legal data, with HR data, to talk to counterparts from outside the enterprise and make sure no sensitive data will be leaked.” The bet is that a neutral broker, owned by neither the model vendor nor the customer, is the only party a risk officer will trust to stand between an agent and the crown-jewel data.

Agent Security Became a Multibillion-Dollar Funding Race

Archestra is one small entry on a very crowded scoreboard. By a tally of agentic-security funding compiled by the analysis site Software Strategies Blog, the top 10 startups in the category have raised a combined $3.6 billion, with the data-security firm Cyera alone accounting for $1.7 billion and the identity vendor Saviynt around $1 billion. In the two weeks around the RSA Conference (RSAC) in March, roughly $392 million in fresh agentic-security funding was announced.

Pure-play MCP security is younger and far smaller, with about $40 million disclosed across four startups. The spread of recent rounds shows how differently the market values the various slices of this problem:

The protocol underneath all of it went from niche to standard inside a year; Anthropic counted more than 10,000 active public MCP servers by December 2025. Incumbents have noticed. The identity giant Okta made its identity platform built for AI agents generally available on April 30, 2026. London keeps feeding the pipeline too, with rounds like a $10 million pre-seed for defence-focused AI landing in the same window.

Where a Seven-Person Team Could Stumble

The momentum is real, and so is the exposure. Seven people and $13.5 million are competing in a category where rivals are raising nine-figure rounds and shipping with hundreds of engineers. Scale is not on Archestra’s side.

The sharper risk is that the independence pitch gets eaten from above. The model vendors Kukuy is betting against, Anthropic, OpenAI and Google, all back MCP and can add native guardrails to their own platforms, narrowing the gap a neutral broker fills. Okta, Oasis and Cyera are pushing into agent governance with budgets and enterprise sales forces an open-source startup cannot match. And open source cuts both ways: 3,700 GitHub stars buy credibility, but turning a free orchestrator into recurring revenue from regulated buyers is a separate, unfinished job, and one the funding is explicitly meant to solve.

The counterweight is the conflict-of-interest argument, and it is a genuine one. A bank’s risk committee may simply refuse to let the company that supplies its model also audit what that model’s agents touch. If that instinct holds across enough of the Fortune 500, a small independent broker has a durable reason to exist no matter how much the giants spend.

For now the company has four Fortune 500 deployments and seven people to run them; the next valuation will turn on whether that ratio holds as the giants move in.