Neurwerk Platform v0.3.29

v0.3.29

• Group gateway bindings by provider category, with admitted API and MCP operations for Studio's developer reference.
• Allow browser requests from the exact configured Studio HTTPS origin on external gateway routes without relaxing gateway authentication.
• Adopt verified Studio 0.20.3 API/Web, extProc 0.19.17, API-key bridge 0.10.4 and Tooling 0.12.0 images, including the Langfuse cleanup image; restore MCP DELETE authentication without changing access rules.
• Move public MCP endpoints to /mcp/<server-id> without old-path aliases; coordinate external clients and the matching LibreChat 0.1.6 add-on revision.
• Disable optional Langfuse product telemetry by default without disabling application traces.
• Before reconciliation, follow the upgrade checklist for grouped-only configuration, ordered MCP metadata migration and retained Valkey storage.

Neurwerk Platform v0.3.28

v0.3.28

• Make Fluent Bit bulk retries idempotent with bounded responses, and recognize native severity prefixes.
• Adopt verified Studio 0.19.8 API and Web images.
• Adopt verified extProc 0.19.16 for stable timestamps in local Ollama chat streams, including the Langfuse cleanup image pin.

Neurwerk Platform v0.3.27

v0.3.27

• Add binding-based AI Gateway routes, inspected proxy trust, personal-provider and Codex subscription access, with explicit user permissions and no fallback to company billing.
• Adopt Studio 0.19.7, extProc 0.19.14, PII Engine 0.15.3, API-key bridge 0.10.1 and Tooling 0.11.2, including current fixes for gateway activity, request IDs, reasoning histories, Responses buffering and error diagnostics.
• Move LibreChat into an independently selected add-on; share generic database and OIDC provisioning while making network access, monitoring and maintenance explicit.
• Require standalone setup CLI 0.4.3 at the manifest's exact revision, and explicit gateway bindings and browser-client approvals before application activation.
• Retire gateway rejected-request capture while retaining cleanup of existing debug data; support separate process logging and expanded client error diagnostics.
• Follow the upgrade steps, including the coordinated LibreChat ownership crossover, before reconciliation.

Neurwerk Platform v0.3.26

v0.3.26

• Adopt verified Studio 0.17.1 API/Web images for administrator-owned MCP setup, shared-key management, tool selection, Publish and Disable, including the native Refresh response fix.
• Authorize MCP discovery through mcp-admin and the caller's own provider connection, with a two-minute native grant and separate invocation permissions.
• Deliver isolated shared keys through OpenBao and External Secrets; verify the active credential version and exact native tool membership before Publish succeeds, and preserve Studio selections across later bootstrap runs.
• Use no-surge rolling updates for managed MCP consumers, including the fix for retained rollout settings during the cutover.
• Adopt verified PII Engine 0.14.0 for bounded, process-local NER caching, enabled by default with a 128 MiB accounted-byte budget per process and a fixed 24-hour entry lifetime; current rules and policy still run on every request.
• Require openbao-stack-setup 0.2.27; activating Studio MCP Setup needs a separately approved clean cutover with matching charts and images, and does not import old MCP selections, shared credentials or personal connections. Follow release/migrations/v0.3.26.md before activation.

Neurwerk Platform v0.3.25

v0.3.25

• Adopt the single, digest-pinned PII Engine 0.13.0 image for Engine and model-sync; local recognition runs on CPU and remote inference services own their hardware.
• Add independent CPU rules and explicit local, remote or disabled NER selection, including language-to-model profiles; existing clients retain legacy selection until they opt in.
• Adopt Studio 0.16.3 for named operator MCP discovery, publication status and automatic checks of approved integrations.
• Isolate MCP provider publication failures and verify approved catalog identity; operator discovery remains disabled until explicitly configured.
• Before activation, adopt the compatible Base release, remove obsolete Engine device/accelerator settings and follow the NER and operator-discovery guidance in release/migrations/v0.3.25.md.

Neurwerk Platform v0.3.24

v0.3.24

• Upgrade LibreChat to the digest-pinned upstream v0.8.8 image and add opt-in cost estimates using the client's AgentGateway pricing catalog and model context windows.
• Install ContextForge by default with private TLS and encrypted database-backed OAuth storage; Vault OAuth support awaits upstream changes. Context7, Brave and GitHub presets remain disabled until explicitly selected.
• Require openbao-stack-setup 0.2.26 and ready ContextForge runtime/database Secrets before adoption; existing standalone ContextForge stages need a single-owner handoff to the default Base stages without deleting the database or encryption keys.
• Adopt Studio 0.16.1, extProc 0.16.2 and API-key bridge 0.8.2 for native MCP identity, personal connections, approved-tool catalog and connection checks.
• Adopt PII Engine 0.12.0-cpu with opt-in remote GLiNER and KServe NER; local analysis remains the default. Remote inference requires approved endpoints, narrow network access and model configuration; KServe also requires pinned tokenizer assets.

Neurwerk Platform v0.3.23

v0.3.23

• Prepare the AgentGateway chart's MCP assertion for 2026-07-28 alongside 2025-11-25, retaining the existing default. New protocol traffic still requires a separately published and adopted compatible extProc image; this chart change does not upgrade the processor.
• Adopt extProc 0.16.0 for scoped MCP 2026-07-28 support alongside 2025-11-25, retaining stateless operation, argument inspection and request-local reversal. Multi-round-trip calls, subscriptions and tasks remain unsupported; mirrored argument headers are rejected when PII is enabled.
• Update extProc to 0.16.1 so chat tool-call continuations preserve common provider reasoning metadata.

Neurwerk Platform v0.3.22

v0.3.22

• Remove legacy Dify charts, releases, and namespace from Base. Clients using Dify need a separately reviewed retirement or add-on handoff before adoption; the new add-on is not selected automatically.
• Move Active Directory federation to an optional Keycloak stage. Existing installations need a planned single-owner handoff before changing the platform source.
• Add unselected shared PostgreSQL and Keycloak interfaces for independent add-ons, without enabling an add-on by default.
• Improve LibreChat recovery after Valkey restarts, K3s metrics, logging, and shared-service stability.
• Add private inspection controls and update pinned inspection images.
• Pin verified Tooling 0.7.4 across its Base chart consumers.

Neurwerk Platform v0.3.21

v0.3.21

• Adopt verified Studio 0.13.0 API and Web images with personal LLM and MCP activity, log and notice improvements. Personal activity remains off unless a client enables it.
• Require openbao-stack-setup 0.2.24 before adoption; reconcile the credential catalog before enabling personal activity so Studio receives its separate Langfuse project keys.
• Group MCP traces by session and reduce routine gateway, ClickHouse, and chart-source logs without disabling application traces or warnings.

Neurwerk Platform v0.3.20

v0.3.20

• Lower routine service, storage, gateway, and collector log volume while retaining warnings and errors; pin extProc 0.14.1.
• Disable unused ClickHouse diagnostic tables without changing application traces or Prometheus metrics.
• Preserve Meilisearch storage across chart upgrades.
• Allow an optional add-on source and configurable Flux log levels in client compatibility checks.