Table of Contents

FoundationaLLM 1.0.0 Release Notes

What's New

This is the first 1.0 release of FoundationaLLM. It builds on 0.9.9 with a new tools experience, first-class memory and sandbox compute, zero-touch Kubernetes deployments, a configurable Cosmos DB data plane, and Agent Portal theming and access control:

  • Tools Experience: Agents are assembled from a searchable Tool Library of tool families, built-in catalogs, and MCP tools. Administrators add, copy, and configure tools from the Management Portal with shared family configuration, registered MCP endpoints or inline URLs, Key Vault–backed headers, and live tool discovery.

  • Programmable Tool Calling: A new FoundationaLLMProgrammableToolCallingTool lets the model write Python orchestration code that calls the agent's other tools as functions inside a custom-container code session, so loops, conditionals, and multi-tool aggregation run in a single invocation.

  • Personalized Agent Memory: Durable, UPN-owned user (and optional per-agent) memory is stored by the FoundationaLLM.Context resource provider and exposed through Core API /memory endpoints. Agents consume memory through an out-of-process Memory MCP server (get_memory / append_memory / clear_memory).

  • Sandbox Code Sessions: Context API can provision Kata-isolated, pod-only sandbox sessions on AKS. Sandboxes are first-class FoundationaLLM.Platform/sandboxes resources with UPN ownership, dedicated Contributor and Admin roles, and full management in the Management Portal.

  • Zero-Touch Managed Endpoint Deployments: The Platform resource provider applies FoundationaLLM.Platform/managedEndpointDeployments directly to AKS — images, replicas, environment variables, secrets, hostnames, and Istio/Kubernetes gateways — without manual kubectl steps.

  • FoundationaLLM Platform Themes: A new artifact package distributes the Agent Portal theme catalog (Interwoven, Star, Sunburst) with footer and focus sections. Themes are managed in the Management Portal and loaded dynamically by the Agent Portal.

  • Platform-Database and Cosmos DB Migration: All platform Cosmos DB access now goes through the well-known Platform-Database data source. Partition keys are configurable per container, instance-id participation is optional, and a re-entrant fllm-cdb migration and verification utility supports multi-tenant rollouts.

  • Conversation Tags, Grouping, and Filtering: Conversations can be tagged, grouped by agent, and filtered with Cosmos DB full-text indexes. Delete and batch operations are available through the Conversation resource provider on Core API.

  • Management API MCP Server: The Management API exposes an MCP server so external agents and tools can invoke management operations with OAuth, using a new MCP.Invoke application permission.

  • MCP Client Tool on SDK v2: The MCP client is upgraded to the 2026-07-28 MCP specification. Runtime context headers (context://user_identity, context://conversation_id, context://agent_object_id) are opt-in rather than implicit.

  • FoundationaLLM.User Resource Provider: User profiles are managed as resources (FoundationaLLM.User/userProfiles) with UPN policy, a Management Portal editor, and a Quick Access section.

  • Agent Portal Access Control: Access to the Agent Portal is gated by a Reader role assignment on the AgentPortal app configuration set. User Portal settings are included in the Agent Portal configuration set.

  • Private Networking for Internal Deployments: deploy/internal can provision Public, Private Endpoints, or full Private Networking (private AKS, VPN, jumpbox) through a management-plane orchestrator.

  • Guest and Federated Identity: Entra ID B2B guest UPNs, consumer Gmail federation, and Google Workspace federation are documented and supported, including identity-substitution pattern updates and in-place UPN editing on Dynamic Compute environments.

  • Realtime Speech Telephony: Core API realtime-speech WebSockets accept Agent Access Tokens in addition to Entra ID, enabling Twilio "Call Me" and other service-bridge telephony clients.

Enhancements and Features

What's New in the User Portal and Agent Portal

  • Conversation tags, grouping, and filtering: Conversations can be tagged and grouped by agent. Filtering uses Cosmos DB full-text indexes for scalable search across large conversation sets.

  • Conversation delete via the resource provider: Conversation deletion is exposed through the Conversation resource provider on Core API, including batch operations.

  • Agent Portal themes: The Agent Portal loads themes dynamically from the AgentPortal app configuration set. Packaged themes include Interwoven (default), Star, and Sunburst, with footer and focus sections and accessible brand-wordmark alternatives (FLLM_LOGO_ALT, FLLM_LOGIN_LOGO_ALT).

  • Agent Portal access: Sign-in requires a Reader role assignment on /instances/<instance_id>/providers/FoundationaLLM.Configuration/appConfigurationSets/AgentPortal. Users without that assignment are denied access after authentication.

  • Improved agent welcome message: Agent welcome-message handling is more consistent when starting and resuming conversations.

  • Realtime speech session interrupt: Switching conversations interrupts an in-progress realtime speech session so audio does not continue against the previous conversation.

  • Keyboard navigation: Focus management and keyboard navigation in the User Portal are improved.

  • Upload and URL-rewrite fixes: Failed uploads no longer leave base64 payloads in the composer; false-positive file URL rewrite issues are resolved.

What's New in the Management Portal

  • Tools Experience (#3140, #3151):

    • Searchable Add Tool gallery grouped by tool family, with recommended tools in library groups
    • Tool Library and Tool Families with shared configuration inheritance
    • Copy-tools wizard that preserves source agent configuration
    • MCP connection editor with inline URL or registered endpoint modes, convert-on-save, List Tools discovery, and Key Vault–backed HTTP headers
    • Structured mini-editors for knowledge-tool query settings and metadata filters
    • Searchable prompt selection in the tool slot editor
    • Agents Contributors can read tools, tool sets, and tool families
  • Themes management: Create, edit, delete, and set the default Agent Portal theme. The branding editor and platform themes package processing (including manifest merge that preserves custom themes) are managed from Platform → Themes.

  • Sandbox management (#3158, #3167): List, filter, create, and delete sandboxes; execute commands; show progress when deleting sandboxes or managed deployments.

  • Managed endpoint deployments (#3152): Create and edit Kubernetes managed deployments, including environment variables (config:// and secure-config://), multiple hostnames, secret references, and improved name generation.

  • User profiles and Quick Access: The FoundationaLLM.User resource provider is surfaced with a user profile editor, policy management, and a Quick Access section.

  • MCP server management: Clone MCP servers, improve endpoint selection, and discover tools on internal MCP servers. Persist inline MCP tool configuration.

  • Data pipeline runs: Filter runs by UPN, improve user selection, and cancel in-progress runs from the portal.

  • Quota events: Improved reporting of quota events and rate-limit quota activity.

  • Role assignments and check access: Improved display, filtering, and check-access flows, including checking access for specified security principals.

  • Data source editor: Partitioning configuration for Azure Cosmos DB data sources (partitioning_strategy, use_instance_id_in_partition_key).

  • Agent list: Agents are filtered and sorted by display name; landing-page and resource-list layouts are improved.

  • Dynamic Compute: In-place UPN editing on Dynamic Compute environments (#3166).

  • Analytics charts: Fixes for analytics display after Cosmos DB changes.

  • Tool editor polish: Clear AI model selection when appropriate; improved tool-name suggestions.

What's New in Agents and Orchestration

  • Programmable tool calling (#3153):

    • New FoundationaLLMProgrammableToolCallingTool
    • Requires a custom-container code session (AzureContainerAppsCustomContainer, DedicatedCustomContainer, or LocalCustomContainer)
    • Scoped, batched execution with progress reporting and a restricted allow-list of callable tools
    • The default Azure Container Apps Dynamic Sessions Code Interpreter backend is not supported (no network egress)
  • Personalized agent memory (#3129):

    • Core API /instances/{instanceId}/memory endpoints for get, append, clear, export, and import
    • FoundationaLLM.Context/memories resources with UPN ownership
    • Memory MCP tool mapping and API reference for MCP server developers
    • context://agent_object_id runtime token so per-agent memory is scoped correctly (#3171)
  • MCP Client Tool:

    • Migrated to MCP Python SDK v2.0.0 (2026-07-28 spec) (#3165)
    • Opt-in system headers via context://user_identity, context://conversation_id, and context://agent_object_id
    • forward_user_identity is no longer honored; migrate headers in the tool editor and re-save
    • File support, improved discovery logging, and inline-endpoint reliability fixes
    • Built-in catalog argument definitions aligned with Genesis, Cursor, PostgreSQL, and Perplexity tools/list schemas
  • Self-service tool sets: Agent templates can set the active self-service tool set. Agents Contributors can read tools, tool sets, and tool families.

  • Workflow status messages: Optional App Configuration lists (WorkflowMainStatusMessages, WorkflowToolStatusMessages) randomize progress messages and support {current_round} / {max_rounds} placeholders.

  • Completion reliability: Improved retries, status messages, and completion-request logging in Orchestration API.

  • Content artifacts: Clients can update content-artifact metadata through Core API.

  • Realtime speech telephony (#3157): WebSocket endpoints accept Agent Access Tokens (X-AGENT-ACCESS-TOKEN) for Twilio Call Me and other service-bridge clients. Agent-name display in realtime speech conversations is corrected.

What's New in Compute, Sandboxes, and Managed Deployments

  • AKS sandbox sessions:

    • environment_type: Sandbox on Context API code sessions
    • FoundationaLLM:APIEndpoints:ContextAPI:Configuration:SandboxCodeSessionService:AKSSandboxDefaults JSON defaults (image, CPU/memory, node pool, gateway, TLS, path rewrite)
    • Service links disabled by default on sandbox pods
    • Command execution and deletion by sandbox id or conversation id
    • Sandboxes Contributor and Sandboxes Admin built-in roles; UPN policy on FoundationaLLM.Platform/sandboxes
    • Context API managed identity needs managedEndpointDeployments/write plus AKS Cluster User and RBAC Cluster Admin on the target cluster
    • New Managed Deployments Admin role (52193588-ce65-4415-a89d-38399ed481e6)
  • Managed endpoint deployments (#3152):

    • KubernetesEndpointDeployment resources applied by the Management API
    • Multiple hostnames on Kubernetes gateways; Istio gateway support
    • HTTPS redirect and managed DNS hostname fixes
    • Environment-variable and secret-reference editing
  • Management API AKS permissions: The Management API managed identity must have AKS Cluster User, AKS RBAC Cluster Admin, and Contributor on resource groups that hold referenced managed identities.

What's New in Data Pipelines

  • Conversation purging (#3155): Optimized conversation-purging plugins and export paths.

  • Conversation Messages data source: Token-usage records can be exported with the conversation messages plugin.

  • Run management: Filter data pipeline runs by UPN; cancel runs from the Management Portal; improved run canonical-id calculation.

  • Editor and validation: Improved data pipeline editor validation, plugin selection, and trigger reliability.

  • Agent tasks: Fixes for agent-task data pipeline execution and timezone handling (full ICU in the Core API Alpine container).

  • Knowledge graph:

    • Publishing stage no longer wipes a published graph on partial or empty runs (#3142)
    • Knowledge-graph queries use the KG vector database rather than the text vector database (#3156)
    • Default vector-database metadata properties include UniqueId and ItemType (#3139)
    • Azure.AI.OpenAI bumped to 2.2.0-beta.4 in Gateway to restore knowledge extraction (#3154)

What's New in Platform and Resource Providers

  • Platform-Database data source (#3163):

    • FoundationaLLM:Instance:PlatformDataSource (default Platform-Database) is the single Cosmos DB configuration source
    • Per-container partitioning_strategy and use_instance_id_in_partition_key
    • Legacy per-API Cosmos DB App Configuration settings are removed and can be deleted after the data source exists
    • fllm-cdb db migrate and fllm-cdb db verify for re-entrant copy, instance-id backfill, guest-UPN mapping, and cutover verification
  • FoundationaLLM.User resource provider: userProfiles with UPN policy, authorizable read/write actions, and Management Portal CRUD.

  • Conversation resource provider: Tags, grouping by agent, full-text filtering, delete, and a batch endpoint on Core API.

  • Authorization:

    • Management API authentication aligned with Core API (RequireScopes, AllowACLAuthorization)
    • MCP.Invoke application permission on Core API and Management API app registrations
    • Core API access tokens include the upn optional claim with externally authenticated UPNs for B2B guests
    • Identity-substitution UPN pattern updated for member and #EXT# guest UPNs
    • Role-assignment matching and Authorization API client-cache race conditions fixed
  • API Endpoint Configurations: FoundationaLLM provider category; MCPServer added to APIEndpointCategory; standalone HTTP headers table on endpoint resources.

  • Artifact package links: Artifact-package provenance is traced on managed resources.

What's New in Deployment and Operations

  • Internal management-plane orchestrator (#3159):

    • Deploy-ManagementPlane.ps1 and master.yaml flow for Entra, resource groups, managed identities, Container Apps environment, Authorization, Management API, and portal
    • Network profiles: Public, PrivateEndpoints, and PrivateNetworking (private AKS, Cosmos lockdown, P2S VPN, Windows jumpbox)
    • deploy/internal is the single source of truth; legacy deployment folders were removed
  • Guest and federated identity documentation:

    • Inviting Entra ID guest users (#3160)
    • Consumer Gmail federation (#3162)
    • Google Workspace federation (#3164)
  • Developer dashboard: Multiple-repo support, service-version rollback, improved deployment status, and local-service management (Restart action, expanded service registry, Linux/macOS HTTP URL overrides).

  • Workload identity: Enable workload identity and OIDC issuer on AKS (see devtools/scripts/k8s/enable-workload-identity.ps1).

What's New in QA and Verification

  • Additional verification suites covering the Tools Experience (V010–V012), Memory MCP (V018), and Genesis MCP smoke tests (V019), plus updates to existing agent-run and scheduled-task verifications.

Improvements

Tools Experience Reliability

  • Legacy agent tool configurations are preserved on save until they are migrated in the editor.
  • Inline MCP URLs display correctly for existing tools; converting an inline URL creates a registered endpoint immediately.
  • MCP HTTP headers persist on save/reload; the unmask toggle no longer resets the value.
  • Tool-family save authorization works for Agents Contributors and built-in catalog overrides.
  • Add Tool gallery search covers the full library; copy-tools no longer drops source configuration.
  • Built-in tool library catalog static initialization no longer returns HTTP 500.
  • Plugin tools are lazy-loaded on first use; AgentTool property reads handle null and native CLR values.

MCP Client Reliability

  • Accidental debug instrumentation that broke plugin load is removed.
  • STDIO upload temporary files are cleaned up; HTTP upload preserves path arguments.
  • Context download no longer assumes a missing user identity.
  • notifications and discovery flows honor registered-endpoint auth headers.
  • MCP SDK v2 transport changes (streamable_http_client, header client, float timeouts) are covered by unit tests and validated against existing Genesis MCP servers.

Sandbox and Managed Deployment Reliability

  • Early guards reject sandbox execute/delete calls that omit ConversationId when the sandbox id is not provided (#3168, #3169).
  • Sandbox filtering can be scoped by conversation id.
  • HTTPS redirect and managed DNS hostname listening issues on Kubernetes gateways are resolved.
  • AKS authentication in the Platform resource provider is more reliable.

Cosmos DB and Analytics Reliability

  • Partition-key creation is centralized; deployment scripts honor use_instance_id_in_partition_key.
  • Migration utility retries, reporting, subscription management, and verification are improved.
  • External-user UPNs are mapped during migration; instance id is backfilled.
  • Analytics charts render correctly after the Cosmos DB configuration change.
  • CosmosClientProvider robustness is increased.

Conversation and Quota Reliability

  • Conversation grouping-by-agent and Cosmos DB filter queries are corrected.
  • Conversation ordering in Core API is fixed.
  • Quota event and rate-limit reporting no longer drop or double-count events.

Identity and Authorization Reliability

  • Guest-account UPN detection regressions are fixed.
  • Authorization service client-cache race condition is resolved.
  • Role-assignment matching in the Authorization API is corrected.
  • Managed-identity authentication to Core API is fixed.
  • Management API MCP OAuth authentication is corrected, including Docker image issues.

Data Pipeline and Knowledge Graph Reliability

  • Data Pipeline API dependency injection and editor issues are fixed.
  • Agent-task pipeline execution and Alpine timezone (ICU) support are corrected.
  • Knowledge-graph publishing no longer clears the published graph on empty or partial runs.
  • Knowledge-graph entity matching uses the correct vector database.

Portal and Developer Experience

  • Management Portal racing conditions that caused instability are resolved.
  • Prompt-list search no longer gets stuck; LangGraph prompts remain in configurable state.
  • File-upload failure no longer keeps base64 content in the composer.
  • Local development: SSL validation skipped for localhost, HTTP ports for Linux/macOS URL overrides, Code Interpreter connectivity to local PythonCodeSessionsAPI, and a Restart action on Manage-LocalServices.ps1.

Pull Requests Included in This Release

The following pull requests were merged to main after the 0.9.9 release (tag 0.9.9). There is no top-level 1.0.0-rc108 tag; this document covers the platform at the *-1.0.0-rc108 service tags (through management-api-1.0.0-rc108).

PR Title Merged
#3150 Cosmos DB optimization 2026-06-08
#3140 Tools Experience UI with endpoint headers, auth fixes, and MCPServer enum 2026-06-15
#3151 Tool Library and MCP tool fixes 2026-06-29
#3152 Zero-touch managed endpoint deployments for AKS 2026-06-30
#3154 Bump Azure.AI.OpenAI to 2.2.0-beta.4 in Gateway to fix knowledge-graph extraction 2026-07-06
#3153 Programmable tool calling 2026-07-07
#3155 Conversation purging improvements 2026-07-08
#3156 Fix knowledge-graph query using the wrong vector database 2026-07-09
#3158 Filtering for sandboxes 2026-07-22
#3129 Personalized agent memory 2026-07-22
#3157 Realtime speech telephony (Twilio Call Me) 2026-07-23
#3159 Internal deployment: management-plane orchestrator and private networking 2026-07-24
#3160 Document inviting guest users 2026-07-29
#3162 Document consumer Gmail federation through Entra 2026-07-29
#3166 In-place UPN editing on Dynamic Compute environments 2026-07-29
#3163 Prepare Cosmos DB features for multi-tenant rollouts 2026-07-30
#3164 Document Google Workspace federation through Entra 2026-07-30
#3165 Migrate MCP client tool to MCP SDK v2 (2026-07-28 spec) 2026-07-30
#3142 Fix KnowledgeGraph publishing stage wiping the published graph on partial or empty runs 2026-07-30
#3139 Add UniqueId and ItemType to default vector-database metadata properties 2026-07-30
#3167 Improve sandbox management in FoundationaLLM.Platform 2026-07-30
#3168 Guard ExecuteCommandInSandbox when ConversationId is missing 2026-07-30
#3169 Guard DeleteSandbox when ConversationId is missing 2026-07-30
#3171 Resolve context://agent_object_id for MCP tool headers 2026-08-02

In addition to the merged pull requests above, this release includes a substantial number of incremental improvements landed directly on main covering Agent Portal themes and branding, conversation tags and grouping, the Management API MCP server, the User resource provider, quota event reporting, data pipeline run filtering and cancellation, Cosmos DB full-text indexes, developer dashboard multi-repo support, and platform-wide reliability and accessibility refinements.

Breaking Changes

Refer to the Breaking changes document for the complete list. Notable items when upgrading from 0.9.9 include:

  • 1.0.0-beta110:
    • The Management API managed identity must have Azure Kubernetes Service Cluster User Role and Azure Kubernetes Service RBAC Cluster Admin on each AKS cluster referenced by managed endpoint deployments, and Contributor on resource groups that contain referenced managed identities.
  • 1.0.0-beta125:
    • The MCP Client Tool no longer implicitly sends X-USER-IDENTITY. Configure context://user_identity and context://conversation_id (and context://agent_object_id when required) as HTTP headers. Re-save tools that used forward_user_identity.
  • 1.0.0-beta126:
    • Create FoundationaLLM:Configuration:AppConfigurationConnectionString pointing to the Key Vault secret foundationallm-appconfiguration-connectionstring.
    • Add a default semantic configuration with a Text content field to the agent-private-store-files and conversation-files Azure AI Search indexes.
    • Enable workload identity and OIDC issuer on the AKS cluster.
  • 1.0.0-rc102:
    • Cosmos DB configuration moves to the Platform-Database data source. Remove the legacy per-API Cosmos DB App Configuration settings after the data source exists.
    • Optional partitioning_strategy and use_instance_id_in_partition_key on the Azure Cosmos DB data source; use fllm-cdb db migrate / db verify when changing partition keys (they are immutable on existing containers).
    • Assign the UPN policy to FoundationaLLM.User/userProfiles, FoundationaLLM.Platform/sandboxes, and FoundationaLLM.Context/memories.
    • Create and assign Sandboxes Contributor / Sandboxes Admin (and Managed Deployments Admin where required).
    • Add the MCP.Invoke application permission to the Core API and Management API app registrations and grant it to MCP server identities.
    • Add the upn optional claim (include_externally_authenticated_upn) on Core API access tokens.
    • Align Management API Entra settings (RequireScopes, AllowACLAuthorization) with Core API.
    • Update FoundationaLLM:Instance:IdentitySubstitutionUserPrincipalNamePattern to accept member and B2B guest UPNs.
    • Grant Reader on the AgentPortal app configuration set to every principal that must use the Agent Portal.
    • Deploy the FoundationaLLM Platform Themes artifact package.
    • Context API requires AKSSandboxDefaults and AKS permissions to provision sandbox sessions.
    • Optional LangChain workflow status-message settings: WorkflowMainStatusMessages and WorkflowToolStatusMessages.

Also review the late 0.9.9 breaking-change entries that may not be present on every 0.9.9 deployment, including the Users Cosmos DB container and Core Worker Reader assignment described under 0.9.9-rc170.

Contact Information

For support and further inquiries regarding this release, please reach out to us:

Conclusion

We hope you enjoy the new features and improvements in FoundationaLLM version 1.0.0. Your feedback continues to be instrumental in driving our product forward. Thank you for your continued support.