Back to vendors

Runner H

Visit site

Browser / computer-use agentindependentVerified 2026-07-01

Autonomous web agent and orchestration platform from H Company that drives browsers with a vision language model, coordinating specialized sub agents to complete multi step tasks from natural language.

Runner H is the flagship agent from H Company, a European artificial intelligence research and product company based in Paris and London, founded in late 2023 by Charles Kantor and Laurent Sifre, a former leading scientist at DeepMind. Backed by a landmark two hundred twenty million dollar round from investors including Accel, Eric Schmidt, Amazon, and UiPath, H builds action oriented agents that execute tasks rather than just answer questions. Runner H is a cloud based, orchestrating agent that turns natural language commands into completed multi step workflows across the web, document systems, and business applications.

The technical foundation is H Company's own family of Holo vision language models, which it has open sourced on Hugging Face. Rather than relying on brittle selectors, the agent reads a screenshot and predicts the precise coordinates to click, interacting with any interface the way a person does and self healing when a layout changes. This computer use approach is state of the art: paired with the Surfer H web agent, H reports a 92.2 percent success rate on the WebVoyager benchmark, ahead of offerings from OpenAI, Anthropic, and Google, while using smaller, specialized, and cheaper models.

Runner H acts as an orchestrating layer that dynamically coordinates specialized sub agents, including the Surfer H browser agent and the Tester H software testing agent, through a reasoning engine. It integrates with existing enterprise applications, supports the Model Context Protocol, and plugs into tools like Google Drive, Notion, and Slack. Developers reach it through an application programming interface for scaled deployment, and individuals can use a browser extension, with a centralized H Studio for designing, monitoring, and debugging automations.

H positions Runner H for enterprise use with GDPR first data handling reflecting its European base, enterprise grade security standards, and privacy technology gained through acquiring Mithril Security. Agents operate under human oversight, and a validator module checks the correctness of results. The clearest gaps against the broadest platforms are the absence of a documented persistent agent memory and of a public marketplace of prebuilt agents, since the model stack is H's own rather than a multi provider gateway. Its strongest suit is best in class autonomous computer use grounded in proprietary, open models.

Vendor details

Canonical URL

https://hcompany.ai

Company status

independent

Use cases & customers

In practice

A finance team faces slow, error prone onboarding. Runner H pre fills multi step verification forms across portals, handles document uploads, and completes the workflow autonomously by seeing and clicking each interface element, freeing analysts for higher value review.

A recruiter needs candidates sourced and logged. From a single natural language prompt, Runner H aggregates resumes from the web, navigates the applicant tracking system, and updates the customer relationship management records without any brittle scripts to maintain.

A quality assurance team must test an e-commerce funnel across payment methods. Tester H converts plain English test scenarios into executable checks, simulates real user actions, and adapts automatically when the interface changes, cutting testing maintenance.

Sources & related URLs

Research notes

Flagship agent from H COMPANY (hcompany.ai). Paris + London. Founded late 2023 by Charles Kantor (CEO) + Laurent Sifre (CTO, ex-DeepMind, AlphaGo/AlphaZero co-author). $220M seed (Accel, Eric Schmidt, Yuri Milner, Xavier Niel, Bernard Arnault, Amazon, UiPath, Samsung, Bpifrance). Team ~70. Acquired Mithril Security (privacy/security). PRODUCTS: Runner H (flagship orchestrating agent), Surfer H (web browsing agent, used by Runner H), Tester H (automated software testing/QA), Holo-1/Holo One (open-source VLMs on Hugging Face + Nvidia NIM + AWS). VLM-BASED: 'sees' screen + predicts mouse-click coordinates from screenshot + NL instruction (Screenspot SOTA), clicks/scrolls like human, SELF-HEALING (adapts to UI changes, no script rewrites). SOTA WebVoyager 92.2% (via Surfer H, beats OpenAI/Anthropic/Google), 5.5x cheaper (smaller specialized models). Runner H = ORCHESTRATING agent (Kantor) — dynamically coordinates specialized SUB-AGENTS, multi-agent architecture (Smart Reasoning Engine + policy/localizer/validator modules), plugs into Google Drive/Notion/Slack, combines multiple agentic models. 4 FULLS: Int=F (MCP + Drive/Notion/Slack + web + API + enterprise-systems), Orch=F (orchestrating-agent coordinates sub-agents autonomously, multi-step end-to-end), Ext=F (Runner-H API + MCP-support + open-source Holo models on HuggingFace, fine-tune/extend), Comp=F (VLM vision-driven browser nav, SOTA WebVoyager 92.2%, Screenspot-SOTA). 8 P's: Know=P (data-extraction + decision-intelligence + validator), HITL=P ('under human oversight' + validator-module assesses correctness), Sec=P (GDPR-first + enterprise-grade-security + Mithril-security-acquisition; NO SOC 2/RBAC named), Obs=P (H-Studio: design/monitoring/debugging), Dep=P (cloud + browser-extension + GDPR-EU-residency + open-source Holo models via Nvidia-NIM/AWS; no full self-host-platform), Trig=P (NL + API + 3rd-party-triggers), Model=P (OWN Holo VLM/LLM stack, open-sourced + fine-tunable + combines-multiple-agentic-models; not multi-provider-gateway), Eval=P (validator-module + dedicated Tester-H QA-agent + SOTA-benchmarks). 2 N's: Mem=N (repeatable-routines but no persistent-agent-memory), Pack=N (specialized sub-agents Surfer/Tester, no user-facing prebuilt-agent marketplace). Use cases: web-testing, BPO (dental-insurance billing/claims), automated-hiring (resume+CRM), QA (e-commerce funnel), financial-onboarding, data-engineering, decision-intelligence. Pricing: free-limited (consumer) + paid + enterprise; private-beta→public-in-waves → self_serve; cost-efficient (5.5x cheaper). Domain hcompany.ai. Score 8.0 (4F/8P/2N).

Capability coverage

8.0 / 14 capabilities · 57%

Integrations & Tool CallingRunner H integrates with enterprise applications including through Model Context Protocol, connects to tools like Google Drive, Notion, and Slack, and acts across web browsers and business systems, so full.	Full
Workflow OrchestrationRunner H is an orchestrating agent that dynamically coordinates specialized sub agents through a reasoning engine to autonomously execute complex multi step workflows across web, document, and business systems, so full.	Full
Knowledge Grounding & RAGRunner H gathers and extracts structured data from the web to inform tasks and decision intelligence and validates results, real grounding in web content short of a documented citation grounded retrieval system, so partial.	Partial
Human Oversight & GuardrailsRunner H is designed to operate under human oversight with a validator module that assesses the correctness of results, real oversight short of a documented runtime approval and guardrail enforcement engine, so partial.	Partial
Security, Identity & GovernanceRunner H offers GDPR first data handling, enterprise grade security standards, and privacy technology from its Mithril Security acquisition, a solid security posture short of documented named certifications like SOC 2 with role based access, so partial.	Partial
Observability & AuditabilityRunner H provides H Studio to centralize automation design, monitoring, and debugging, real observability short of a documented full per action audit and traceability system, so partial.	Partial
Memory & State PersistenceRunner H supports repeatable routines, but a first class persistent cross session agent memory could not be verified.	Unable to verify
Deployment & Data ResidencyRunner H runs as a cloud platform and browser extension with GDPR first European data handling, and its underlying Holo models are open source and deployable through Nvidia and AWS, real deployment flexibility short of a documented full self hosted agent platform, so partial.	Partial
Prebuilt Agents, Templates & PacksRunner H coordinates specialized agents like Surfer H and Tester H, but a browsable library of prebuilt or cloneable agents for users could not be verified.	Unable to verify
Triggers & Channel CoverageRunner H automations run from natural language, through its application programming interface, and can be triggered from third party systems, real triggering short of broad customer facing channel coverage, so partial.	Partial
Model Flexibility & RoutingRunner H runs on H Company's own Holo vision language and language models, which are open sourced and can be fine tuned, and it combines multiple specialized agentic models, real model capability short of a documented multi provider routing gateway, so partial.	Partial
APIs, SDKs & MCP ExtensibilityRunner H exposes an application programming interface for integration and scaled deployment, supports Model Context Protocol, and open sources its Holo models on Hugging Face for developers to run, fine tune, and extend, so full.	Full
Testing, Debugging & OptimizationRunner H includes a validator module that assesses the correctness of its results, a dedicated Tester H agent for automated software testing, and strong public benchmark performance, real testing and validation short of a documented agent evaluation and optimization harness, so partial.	Partial
Browser & Computer UseRunner H and its Surfer H agent navigate browsers like a human using a vision language model that reads the screen and predicts precise click coordinates, achieving state of the art results on the WebVoyager benchmark, so full.	Full

Recent platform changes

No recent material changes tracked yet.

View all changes for Runner H →

Pricing

Public access in waves; tiered subscription with usage limits, enterprise contact sales

Contact onlyMedium variable cost

Verified 2026-07-01

Data confidence: high