Skip to main content

​
v1.19.36 — May 8, 2026

​
MCP gateway for Claude Code, SCIM provisioning, Context Graph wiki layer, SQL safety tiers, and broad platform reliability improvements

​
MCP Gateway for Claude Code and External Tools

Genesis now exposes a first-class MCP (Model Context Protocol) gateway that lets Claude Code and other MCP-compatible clients connect directly to Genesis agents, datasets, and apps without manual credential setup.
  • A Genesis MCP gateway facade routes MCP tool calls to the correct Genesis agent threads, datasets, and app actions, giving Claude Code and compatible IDEs a single connection point into your Genesis deployment.
  • An npm bridge package for Snowflake SPCS deployments lets external MCP clients reach a Genesis instance running inside SPCS without requiring direct network access — the bridge handles credential handoff and status URL resolution automatically.
  • Hosted gateway threads now auto-start and maintain ownership across MCP sessions so results are always returned to the originating thread.
  • The gateway exposes dataset listing, app helpers, and wiki search as named MCP tools, with formalized response contracts that external clients can rely on.
  • Claude-specific fast paths streamline the most common data-query and app-review workflows when accessed through the MCP protocol.

​
SCIM Provisioning

Genesis now supports inbound SCIM 2.0 provisioning for automated user and group lifecycle management.
  • SCIM provisioning allows your identity provider to create, update, and deactivate Genesis users and group memberships automatically, eliminating manual role assignments at scale.
  • AWS-sourced SCIM group reconciliation was hardened to handle edge cases in large directory environments.
  • Public documentation for SCIM configuration is now available for both EKS and SPCS deployments.

​
Context Graph: Wiki Layer, Expanded Lineage, and New Scanners

The Context Graph gained a semantic wiki layer, three new data source scanners, and a more reliable index rebuild path.
  • A new wiki layer sits over the Context Graph, letting agents and operators annotate any data asset — table, schema, pipeline, or application — with curated knowledge that is retrieved automatically when that asset appears in a conversation.
  • A dedicated Wiki panel in the Context Graph viewer surfaces wiki content alongside lineage edges, with intermediate pages synthesized automatically for hierarchy levels that have no direct annotation.
  • The Context Graph now includes Fivetran as a first-class source: the Fivetran scanner harvests connector-to-destination lineage and links it to existing warehouse assets.
  • dbt Cloud is now a first-class orchestration node type in the graph: dbt model runs appear as dbt:// nodes connected to their upstream sources and downstream warehouse tables.
  • Microsoft SQL Server (MSSQL) can now be scanned for schema lineage through a native connector — no third-party bridge required.
  • GitLab lineage scanning now matches GitHub parity: pipelines, repos, and CI-driven table references are tracked with full edge coloring.
  • Scan cancellation is now reliable across all scanner types — cancelling a long-running scan takes effect promptly.

​
Agent Knowledge (Wiki Learnings) Improvements

The wiki learnings system received major reliability and quality upgrades, with a new dashboard GUI for operators.
  • Operators can now create, edit, and delete wiki learnings directly from the Genesis dashboard without needing an agent conversation.
  • The librarian now detects and corrects superseded or contradictory learnings automatically using two independent correction detectors (supersession-based and LLM-classifier-based), keeping the knowledge base accurate over time.
  • The wiki auto-ingestion loop runs on a configurable periodic schedule, ensuring new assets discovered by scanners are incorporated into learnings without manual triggers.
  • Wiki learnings now carry a validation state and privacy class, giving operators fine-grained control over which learnings are surfaced to end users.
  • A full-text reverse index enables cross-entity discovery: a single learning about a table can now surface when related schemas or dashboards are mentioned in a conversation.
  • Wiki admin inspection tools are available to operators for diagnosing the state of the learnings database.

​
SQL Safety Tiers

Agent-executed SQL statements are now checked against a tiered safety policy.
  • A new CATASTROPHIC tier blocks agents from executing DROP DATABASE, DROP SCHEMA, and equivalently destructive statements — even if the agent explicitly constructs such a query.
  • A secondary intent-match guard further validates that the SQL the agent is about to run matches the user’s stated intent before execution.
  • Safety guard activations are now written to the audit log with user identity and thread context, giving compliance teams a complete record of blocked operations.

​
Access Control and RBAC

RBAC controls were expanded with several new capabilities that give administrators and team leads finer-grained governance.
  • The Edit Agent UI is now gated behind the manage_agents permission, so contributors who should only read agent configs cannot make changes accidentally.
  • Per-role PAT-connection capability is now configurable: administrators can decide which roles are allowed to create and use personal access token connections.
  • Missions and Genesis Apps that belong to a shared role are now visible to all members of that role, enabling shared workspaces across teams.
  • Built-in agent definitions now expose per-block toggle controls and a persona-only editing mode, giving administrators control over which methodological behaviors are active without overwriting all agent instructions.

​
Slack Integration: Secret Forms

Slack-backed conversations can now request sensitive credentials from users through secure, ephemeral form prompts.
  • When an agent running in Slack needs a credential that has not been pre-configured, it can now prompt the user with a structured secret form directly in the Slack thread. The value is captured securely and never exposed in the conversation history.
  • Secret form prompts are automatically cancelled when the user sends a new message or a different user takes over the conversation, preventing stale prompts from accumulating.

​
Connections and AI Providers

  • Databricks OAuth connections are now supported for LLM providers, enabling single-sign-on AI access for Databricks-hosted customers.
  • AI provider LLM model assignments are now reconciled across all configured providers on server startup, so manually configured custom models are visible immediately after a restart.
  • GitLab OAuth credential synthesis now correctly generates usernames for the git credential helper, fixing push and pull failures on GitLab-backed managed repos.
  • GitLab managed repos now match GitHub parity for all create, clone, branch, and delete operations.
  • Snowflake PrivateLink OAuth URLs are now handled correctly alongside standard Snowsight URLs.

​
Genesis Apps Improvements

  • Agents building Genesis Apps are now warned when their edits target the working copy while the user is viewing the promoted production version, preventing confusing mismatches.
  • Viewer tokens are now forwarded automatically for runtime API requests from app embeds, eliminating blank dashboards for users with viewer-only access.
  • The agent is now guided to use edit-in-place operations rather than full rewrites, improving iteration speed and reducing regression risk on large app files.

​
Chat, Threads, and Notification Improvements

  • Interactive credential cards and agent action request prompts in chat are now anchored inline to the message that triggered them, so the context is always visible when filling in a form.
  • All interactive cards in a thread are now automatically cancelled when the user sends their next message, keeping the conversation flowing cleanly.
  • Thread ownership and legacy orphaned thread assignments are now repaired automatically on startup.
  • Dormant thread review is available to help administrators identify and clean up inactive threads.

​
Performance

  • Thread hydration for large conversations is now significantly faster: the server caches pre-formatted message payloads and deduplicates redundant WebSocket subscribe calls, reducing load times for threads with hundreds of messages.
  • The thread-service compression pipeline now uses a bytes-native size target, a parallel-turn protection lock, and an overhead-saturation safety valve to prevent runaway context growth under concurrent load.
  • Context state is now stored and read through the thread service instead of a JSON sidecar file, reducing storage I/O on every turn.
  • Startup time improved: the stage-wide cleanup audit is disabled by default and the thread index is seeded with only the 200 newest threads on first boot.
  • The LLM provider cost calculation now correctly accounts for cached token usage, producing accurate cost estimates on repeated queries.

​
Platform and Deployment

  • Block storage on Snowflake SPCS deployments now migrates stage-mounted data to the block storage volume on startup, so existing deployments can adopt block storage without re-provisioning.
  • Support bundle downloads now scrub schema-sensitive data from tool-call outputs and offer a cascading mail-client delivery path.
  • Agents that soft-delete still retain their thread history, preserving conversation context for administrators and audit purposes.