Owkin

K Pro connects to customer systems registered as MCP servers, Owkin-provided wrappers or the customer's own, naming Benchling and Dotmatics for assay data, Snowflake, S3, Databricks and GCS for queries, KNIME and Pipeline Pilot as pipelines and Spotfire and Tableau for writing outputs, with per-tool authorization through OIDC, SAML, service accounts or API keys.

Sourcedocs.owkin.com/connect-and-integrate/mcp-and-agent-toolsread 2026-09-29

One orchestrating agent routes each request to specialized agents (Multi-Omics, Consensus, Literature Review, Prior Knowledge, Pathology Explorer, Plotter, Code Execution, Discoverability), runs parallel autonomous sub-agents on different angles, and follows skills that encode multi-step workflows.

Sourcedocs.owkin.com/what-you-can-do-with-k-pro/how-k-pro-worksread 2026-09-29

A data catalog Owkin maintains, with Owkin managed datasets including MOSAIC spatial and multiomics data, a 104 hospital patient data network and 19 public datasets, joins the customer's own data, which can stay in the customer's infrastructure, move into an Owkin managed environment or be queried in Databricks or Snowflake. Retrieval augmented generation checks that cited sources are genuine.

Sourcedocs.owkin.com/integrate-your-data/connecting-your-data-sourcesread 2026-09-29

Agentic behavior is bounded by role based permissions, per-tool authorization and safety checks, and custom skills pass a publish-and-review lifecycle before use. Customers set these limits themselves, and no approval step before an agent's action takes effect is described.

Sourceowkin.com/k-os/k-pro-faqread 2026-09-29

Sign in runs through SSO over SAML 2.0 and OIDC with Okta, Entra ID and Cognito, and role based permissions and least privilege roles are enforced across direct calls and calls made through MCP. Owkin holds ISO 27001:2022 (and ISO 13485:2016) and runs a Vanta trust center.

Sourcedocs.owkin.com/connect-and-integrate/sso-and-authenticationread 2026-09-29

Each run shows the prompt and session context, the tools invoked and data services queried, and a step by step orchestration trace. All events go to an audit trail with request tracing from end to end, LLM usage is logged and the order of tool calls can be traced, with records kept separate for each customer.

Sourcedocs.owkin.com/governance-and-security/model-governance/explainability-and-traceabilityread 2026-09-29

Chat history is saved per user and organization and project workspaces persist, which carries conversation state forward. No other memory that agents keep between conversations is described, and skills are instructions people write.

Sourceowkin.com/k-os/k-pro-faqread 2026-09-29

K Pro deploys as an Owkin hosted, single tenant SaaS or in a dedicated cloud environment in the customer's own AWS, Azure or GCP organization (BYOC, with a least privilege role across accounts), and the Professional tier offers on premise or client instance deployment behind the customer's firewall. Hosting regions are not named.

Sourcedocs.owkin.com/governance-and-security/infrastructure-and-hostingread 2026-09-29

Skills certified by Owkin and skills drawn from key opinion leaders (KOLs) ship as prepackaged workflows of several steps, beside eight named specialized agents, each with its own job. A use cases library covers target identification, indication prioritization, drug positioning, and literature and patent review.

Sourcedocs.owkin.com/what-you-can-do-with-k-pro/how-k-pro-worksread 2026-09-29

Work starts when a scientist asks in chat, when a job is submitted through the REST API, or when K Pro is called over MCP from the user's own systems. K Pro is not designed to start actions on its own and cannot operate without authenticated users, so nothing starts work on a schedule or event.

Sourceowkin.com/k-os/k-pro-faqread 2026-09-29

K Pro reasons through an enterprise LLM that can be substituted according to need, beside Owkin's own OwkinZero and Pathology Explorer models. No provider is named and who makes the substitution is not stated.

Sourceowkin.com/k-os/k-pro-architectureread 2026-09-29

A public REST API for K Pro supports job submission, status polling and result retrieval as structured JSON with provenance, authenticated by API keys or service accounts, and MCP integration lets K Pro be called from a user's existing internal systems. The list of endpoints is not published.

Sourcedocs.owkin.com/connect-and-integrate/api-documentationread 2026-09-29

Skills carry success criteria with a quality threshold, validation and reproducibility checks and an example execution, and output must pass schema validation. Custom skills are versioned and promoted through a publish and review process, so each workflow the agents follow is checked before it goes into use.

Sourcedocs.owkin.com/what-you-can-do-with-k-pro/writing-reproducible-skillsread 2026-09-29

No browser, desktop or computer control is described, and K Pro runs the code it generates in a sandbox.

Sourceowkin.com/k-os/k-pro-architectureread 2026-09-29