Atlan

Most of what the context agents do lands inside Atlan: descriptions, READMEs, glossary terms and term links written onto catalog assets. Outside it, Context Engineering Studio deploys a context repository into Snowflake as a semantic view, and into Databricks by pushing descriptions to Unity Catalog, creating metric views and configuring a Genie space, under permissions granted at deploy time.

Conversational AI answers in Slack and Microsoft Teams, and agent templates read GitHub and Confluence through a read only credential and take skills of the customer's own, which are instructions rather than tools. The MCP gateway that governs agent access to outside tools belongs to Atlan Trident, a separate product in private preview; opened to every tenant, it would let registered agents act in those tools under Atlan's policies.

Sourcedocs.atlan.com/product/capabilities/governance/context-engineering-studio/concepts/what-is-context-engineering-studioread 2026-11-09

Deep mode in conversational AI runs an agent loop that reasons through a question in as many steps as it needs, pulling lineage, glossary and asset context and, when asked, writing descriptions as part of the answer.

Context agents work through a collection in stages: a sample run that writes nothing, a confirmation showing sources, scope and instructions, then a complete run that writes to the glossary, with one agent running at a time per tenant. Context Engineering Studio chains asset proposal, approval, model assembly and deployment. Customers cannot compose their own multiagent graphs or hand work from one agent to another.

Sourcedocs.atlan.com/get-started/concepts/how-credits-workread 2026-11-09

The Enterprise Data Graph keeps a persistent, queryable model of the customer's estate: metadata from more than 80 connectors, column level lineage, glossary terms, domains, data products, quality rules and certificates, refreshed by connector workflows.

Semantic search runs over embeddings held in a regional vector store, so agents can ask across the catalog in natural language.

Context repositories add a governed semantic model per use case with verified question and SQL pairs and custom instructions, and knowledge folders accept uploaded documents that are indexed for retrieval, a feature still in early preview.

Sourcedocs.atlan.com/product/capabilities/atlan-ai/concepts/securityread 2026-11-09

Every MCP tool that changes something in Atlan proposes first, returning a preview of the change and waiting for the user's approval before it writes.

Admins decide in Labs who can use MCP and, separately, who can create and update metadata from conversational AI, scoped to admins, selected personas or everyone, and support can switch a tenant to read only mode.

The term linker writes only links from a sample a person reviewed, agent output is tagged as AI generated while human written text is left alone, and governance workflows route metadata changes through multistep approvals to their owners.

Sourcedocs.atlan.com/product/capabilities/atlan-ai/references/mcp-toolsread 2026-11-09

Sign in runs through SAML 2.0 single sign on, with setup guides for Okta, Microsoft Entra ID, Google, JumpCloud, OneLogin and PingFederate, and SCIM provisioning from Okta or Azure AD syncs users and groups. Admin, Member and Guest roles, admin sub roles, personas, purposes and metadata, data, AI and domain policies decide who sees and changes what, and every MCP call inherits them, with service accounts limited to exactly one persona. Atlan holds SOC 2 Type II, ISO 27001 and ISO/IEC 42001, attests to HIPAA through an external auditor's letter, and supports customer managed encryption keys.

Sourcedocs.atlan.com/faq/security-and-complianceread 2026-11-09

Each tool call across MCP, conversational AI, Slack and Teams lands as a row in the AI_TOOL_CALLS table of the Lakehouse usage namespace, with the interface, tool name, user, client, auth type, timestamp, duration and error flag, refreshed daily.

The MCP server also logs a structured event per call with the model identifier, request ID, status and originating query, along with denials and authentication failures, and redacts arguments and responses first.

Asset activity history and an audit search API show who changed what and from which client, AI edits carry a marker, and sign in events stream to the customer's own bucket for a SIEM.

Sourcedocs.atlan.com/platform/lakehouse/references/usage-analytics-namespaceread 2026-11-09

Conversational AI keeps a personal memory for each user, holding assets and topics discussed, corrections to earlier answers, stated preferences and work context such as role, team and project names, so follow ups in later sessions carry over. Other users never see it, and sensitive items such as email addresses, phone numbers, government IDs and API keys are discarded even on request. There is no settings screen and no stated retention period; a user removes an item by asking the assistant to forget it. Memory works only in the Atlan app, not in Slack, Teams or the browser extension.

Sourcedocs.atlan.com/product/capabilities/atlan-ai/references/memory-in-conversational-airead 2026-11-09

Tenants run on AWS, Azure or Google Cloud, and the data processing agreement keeps customer data in the hosting region the customer selects, with Google Cloud tenants available in four United States regions and five European ones. The AI gateway operates in US, EU and APAC regions and handles each tenant's requests in its nearest region, with PrivateLink and VPC peering between the tenant and the AI infrastructure. A paid Self Deployed Runtime runs Atlan connector apps inside the customer's own infrastructure, while the catalog itself stays a hosted service in an Atlan managed VPC.

Sourcedocs.atlan.com/platform/concepts/atlan-network-detailsread 2026-11-09

Context agents come as named specialists: Scribe writes table and column descriptions, Doc writes READMEs, Lexis bootstraps a glossary, Nexus links columns to business terms, Sage finds conflicting metric definitions, Atlas tags domains, Vera scores data quality and Orion maps an ontology.

In the catalog, the Description, README and SQL Intelligence agents run on collections, and a skill hub offers ready made skills such as a catalog health report, a governance coverage report, lineage impact analysis and an AI usage report, which a workspace copies and edits. Agent templates in private preview add a glossary term generator and a term linker.

Sourceatlan.com/context-agentsread 2026-11-09

Context agents run on daily, weekly or monthly schedules with a start time and timezone, and keep fresh collections regenerate AI written descriptions and READMEs when an underlying asset changes, processing only what changed since the last run. Conversational AI takes questions in the Atlan app, in an assistant panel on any asset page, in Slack channels, in Microsoft Teams conversations and through a browser extension inside Snowflake, Databricks and Power BI. Outside agents reach the same context over MCP from Claude, ChatGPT, Codex, Cursor, VS Code, Gemini CLI, n8n and Copilot Studio.

Sourcedocs.atlan.com/product/capabilities/governance/context-agents-studio/how-tos/run-and-schedule-an-agentread 2026-11-09

Atlan manages the models behind its own AI centrally, routing requests through a gateway that reaches Anthropic Claude on Microsoft Azure, OpenAI GPT models and open source models Atlan hosts, with a fixed model assigned to conversational AI, its fast mode, context agents and embeddings.

Customers pick an answer mode, Quick, Deep or Adaptive, which changes reasoning depth and credit cost but not the model. Agents reading Atlan over MCP keep whatever model their own client runs.

The Model Gateway that lets an organization connect its own providers and choose models belongs to Atlan Trident, in private preview; opened to every tenant, it would put that choice in the customer's hands.

Sourcedocs.atlan.com/product/capabilities/atlan-ai/concepts/securityread 2026-11-09

A hosted MCP server, on for every tenant at one public endpoint, takes OAuth sign in per user or an API key for service accounts and exposes 41 tools: search, lineage and read only SQL, plus write tools that update descriptions, certificates and READMEs, set custom metadata and tags, and create glossaries, terms, domains, data products and data quality rules. Errors return a structured envelope with a retry flag. Beyond MCP, a REST API, the pyatlan Python SDK, a Java SDK and OAuth clients for machine to machine access cover users, groups, personas, policies, assets and data contracts.

Sourcedocs.atlan.com/product/capabilities/atlan-ai/how-tos/remote-mcp-overviewread 2026-11-09

Context Engineering Studio leaves testing to Atlan Trident, whose evals run an agent against fixed cases with code scorers or hosted judges, store experiments for comparison and can gate pull requests in CI, but Trident is in private preview and its Evals pages need enablement per organization.

Opened to every tenant, those evals would put scored test cases and a CI gate on each agent change. Inside the catalog, agent templates offer sample runs that publish nothing so a team can tune instructions first, and context agents score each output on accuracy, clarity, style and completeness, applying confident ones and sending the rest to reviewers.

Sourcedocs.atlan.com/product/capabilities/governance/context-engineering-studio/concepts/what-is-context-engineering-studioread 2026-11-09

Atlan's agents do not drive a browser, desktop or remote computer. The browser extension brings Atlan context and conversational AI answers into Snowflake, Databricks and Power BI pages for the person reading them, and Atlan Trident Desktop imports local skills and coding sessions into a registry rather than operating applications. Work reaches other systems through connectors, MCP tools, read only SQL against connected warehouses and deployments into Snowflake and Databricks, all programmatic interfaces.

Sourcedocs.atlan.com/product/capabilities/atlan-ai/references/access-conversational-airead 2026-11-09