Modus

**Product and the concrete problem it solves.** Modus is aimed at a specific failure mode in enterprise AI: a model can be connected to a company's warehouse, BI tools, code, documents, and collaboration systems and still not understand what the business means. A finance agent may retrieve several tables that all contain a revenue-like field but fail to know which definition the company trusts; a support agent may find a current policy without the discussion that explains its exception; and an internal analyst may receive a large, noisy context window that increases latency and cost without improving the answer. Modus calls this the Context Gap. Its product, the Context Warehouse, is intended to become a live system of understanding alongside a company's data warehouse. It learns how teams actually work, preserves the definitions and decisions that matter, and gives an agent a focused, permission-checked brief for a particular question or workflow. The practical promise is not another chatbot or data catalog. It is an infrastructure layer that makes existing and future agents more accurate and less wasteful while making their access to enterprise knowledge governable.

**Core technology and how it works.** The public product description outlines four connected stages: Connect, Learn, Govern, and Compose. Modus connects to structured and unstructured systems including data warehouses, BI, dbt, pipelines, code, documentation, and collaboration tools. Its stated metadata-first approach means it learns from how data is organized and used rather than requiring raw customer data to be copied into a separate system. The learning layer continuously surfaces trusted definitions, metric calculations, relationships, and operational conventions from the work already taking place. The governance layer packages permissions, tools, and context into Scopes for a particular team, agent, or workflow and applies those controls before information reaches a model. The composition layer selects and assembles the relevant subset at query time, serving it through an API or Model Context Protocol connection. The company's own site claims up to ten times fewer tokens per query, which is a company-reported efficiency claim rather than an independently audited benchmark. Its technical blog gives a more detailed engineering picture: typed hierarchies for heterogeneous sources, refresh cadences, watermark-based synchronization, multiple retrieval strategies, explicit tenant isolation, ACL checks at the final content boundary, workflow orchestration, execution identities separate from mining identities, and evaluation datasets for measuring retrieval quality. That description is unusually concrete for an early-stage company, although public material does not disclose architecture diagrams, scale limits, model choices, latency distributions, or customer benchmark results.

**Market, customers, and go-to-market.** Modus sells into the infrastructure problem created when enterprises move from isolated AI pilots to agents that operate across business systems. The likely economic buyers are data-platform leaders, AI platform teams, security and governance teams, and application owners who need agents to answer or act reliably without granting them uncontrolled access to an entire data estate. The product can be introduced through a focused use case such as a finance analyst agent, customer-operations workflow, or product-analytics assistant and then expanded as more systems are connected and more Scopes are defined. The company supports both bring-your-own agents, including Claude, Cursor, and in-house systems, and agents or workflows built directly on Modus. That model gives it a chance to become an underlying service rather than a destination application, but it also creates a demanding integration and proof-of-value motion. Modus must show that its context improves answer quality and reduces model consumption enough to justify another platform in an already crowded data stack. The July 2026 coverage says the company had twelve employees, all based in Israel, and customers already in production according to a founder post; public sources do not name those customers, disclose contract sizes, or provide revenue, retention, deployment counts, or sector mix. The official website presents integrations and a demo-led enterprise motion, but does not publish a self-serve price or a public product usage dashboard.

**Traction, funding, and third-party validation.** Modus emerged from stealth in July 2026 with a $10 million Seed round led by Insight Partners, with Soma Capital and a group of technology entrepreneurs and executives participating. CTech identifies participants including founders of Cyera and Epsagon, Eyal Kishon, and Nadav Abrahami of Wix and Dazl. The New Stack independently reported the same funding and described the product as a layer that maps how a business operates across its systems before handing an agent only the relevant slice of that map. Those sources provide credible evidence of a real company, a financed product effort, and a clear category thesis, but the financing should not be mistaken for product-market proof. The public record supports a company-stated claim that customers were already in production at launch, while it does not support a named-customer list, paid conversion rate, annual recurring revenue, customer concentration, or independent accuracy study. The official product site adds important validation of what is actually being offered: MCP and API delivery, SDK and CLI access, context-as-code configuration, exportable context and configuration, permission-scoped workflows, human-in-the-loop controls, and security claims including SOC 2 Type II and ISO 27001. The legal and data-processing pages identify Modus Artificial Intelligence Ltd. in Tel Aviv alongside a Delaware entity and describe a security program with encryption, least-privilege access, penetration testing, and independent attestations available to customers under NDA. The certifications are company-published and should be checked directly in diligence.

**Founders and team background.** Modus was founded in 2026 by Daniel Shimoni, co-founder and CEO, and Tomer Mesika, co-founder and CTO. Their backgrounds are relevant because both came from companies where enterprise data quality and data context are central operating problems. Shimoni previously served as Vice President of Product at Lusha, a go-to-market data platform, while Mesika was Head of Architecture at Cyera, a data-security company. The founders told The New Stack that they left their prior roles in September 2025 after repeatedly encountering the gap between making information available to AI and making that information meaningful enough for reliable work. This is a product-led founding thesis supported by direct experience with data-intensive enterprise systems rather than a generic wrapper around a public model. The public team is still small: CTech reports twelve employees at launch, all in Israel, and the official About page names the two founders but does not publish a larger executive roster. That combination is a strength and a constraint. Shimoni's product background may help translate a technically subtle infrastructure problem into a buyer-facing workflow, while Mesika's enterprise architecture experience is relevant to multi-source synchronization, access controls, and production reliability. Conversely, the company has not publicly demonstrated the sales, customer-success, security-operations, or large-scale platform engineering depth required to become a control layer across sensitive enterprise systems.

**Competitive dynamics.** Modus competes with several overlapping approaches, none of which is identical to its proposed Context Warehouse. Data catalogs and semantic layers such as Atlan, Alation, and dbt's semantic tooling organize definitions and lineage but generally depend more heavily on explicit modeling and do not necessarily compose context for every agent at runtime. Vector databases and retrieval-augmented-generation frameworks provide important indexes but do not by themselves solve source hierarchy, freshness, permission propagation, or the distinction between a search result and a business-approved definition. Data platforms such as Databricks and Snowflake can add AI and governance features from a position of control over core enterprise data, creating a powerful bundling threat. Agent-governance platforms such as Runlayer and Zenity approach the problem from tool access, runtime policy, and agent security rather than from learned business context, while AI application platforms may embed their own retrieval and memory layer. Modus's claimed edge is the combination of continuous observation of how a company works, task-specific context composition, model and warehouse neutrality, and an explicit governance boundary before prompt delivery. The defensibility question is whether the continuously accumulated organization-specific context and permission graph become hard to reproduce, or whether hyperscalers, warehouses, catalogs, and agent platforms can add comparable features as a bundled extension. Its technical blog shows awareness of hard systems issues such as stale context, multi-tenancy, orchestration, ACL propagation, token budgets, and evaluation, but public evidence does not yet demonstrate that Modus handles them better than incumbents at production scale.

**Defense, security, and resilience relevance.** Modus has credible dual-use relevance through information integrity and controlled AI operations, not through a disclosed defense product. Defense organizations, emergency-management bodies, utilities, hospitals, and other critical-infrastructure operators increasingly need software agents to work across fragmented operational, maintenance, logistics, intelligence, and compliance systems without exposing everything to every automated process. A context layer that preserves authoritative definitions, scopes access by mission or workflow, and keeps an audit trail of what an agent was allowed to see can reduce two strategic risks: an agent acting on stale or contradictory information, and an agent receiving sensitive data outside its operational need. Model-neutral API and MCP delivery also fit environments that must change models or host systems without rebuilding their information layer. The company’s Israeli legal entity and the founders’ experience at Lusha and Cyera make the commercial-to-security translation plausible, and the product's security controls are relevant to regulated deployments. The calibration is important. Modus has not publicly disclosed an Israeli Ministry of Defense, military, intelligence, public-safety, utility, or defense-industrial customer; it has not shown operation in disconnected or degraded networks; and it has not published defense accreditation, classified-environment support, sovereign hosting, or mission-specific evaluations. The strategic case is therefore resilience optionality around governed enterprise AI, with no evidence yet of a fielded national-security capability.

**Growth stage, trajectory, and diligence risks.** Modus is classified as early: it was founded in 2026, raised a Seed round shortly after formation, employs twelve people according to launch coverage, and has a public product with production customers claimed but not named. The trajectory depends on becoming a shared understanding layer used by many agents rather than a narrow retrieval product attached to one workflow. The upside is significant if the Context Warehouse becomes the place where enterprise definitions, permissions, corrections, and operational memory accumulate and can be reused across models and applications. The central diligence risks are specific. (1) Product proof: the tenfold token-reduction claim and the accuracy benefit need independent, task-level comparisons against strong retrieval and semantic-layer baselines. (2) Data and privacy: Modus's marketing emphasizes metadata-first processing, while the legal pages also say that relevant data may be submitted to third-party AI providers for core features; customers need a precise data-flow, retention, tenancy, and model-training assessment. (3) Permission correctness: a single stale ACL, wrong tenant boundary, or incorrect inferred definition could create a security or operational incident, especially when agents can act rather than merely answer. (4) Adoption friction: integration across data warehouses, BI, code, documents, and collaboration tools can become a long platform project before a buyer sees measurable value. (5) Competitive response: Databricks, Snowflake, Microsoft, cloud providers, catalogs, and agent-security vendors can bundle context and policy features into existing contracts. (6) Commercial maturity: named production customers, revenue, retention, deployment scale, and renewal evidence are not public. Milestones to watch are independent retrieval-quality results, larger named deployments, clear evidence of reduced inference cost, audited security documentation, support for private or restricted environments, and proof that customers reuse the same context across multiple agent workflows.