What changed, and when

Google Gemini on Vertex AI is now available as a chat model provider, and usage metering now counts cached and reasoning tokens correctly.

The MCP server has a published guide for deploying it on Google Cloud Run, and it now honours the PORT setting Cloud Run provides.

The pipeline builder gains a step-type palette and a rule-based routing step.

The MCP relay is published as version 0.1.1 with a container image and a health check.

Platform-operator tools are now separate from tenant administration, operator sign-in adds multi-factor authentication, and a personal organization is created at first sign-in.

Slack and email notifications are sent when an investigation run completes.

The plugin repository can be installed as its own Claude Code plugin marketplace.

Model spend is rolled up in the web app, and the run overview shows a verdict for each hypothesis.

Approval grants carry a risk ceiling, so a time-boxed grant pre-approves low-risk actions only by default (Beta).

Settings gain a policy section and webhook trigger configuration, a built-in pipeline catalog is included, and a run now tells you when its budget cut an investigation short.

A live stream of the agent's reasoning shows a summary of each step as the investigation runs.

Enterprise administrators can set allowlists for inbound MCP callers and for their organization's own model and MCP use.

Self-hosted installs get runtime configuration and a first-admin bootstrap step. Self-hosting is available in closed beta.

Evidence is stored durably, with a retention period you can set per tenant.

The pmctl operator command line (login, start, status, report, events, approve, reject) is published on npm as production-master.

The Python SDK is published on PyPI as production-master 0.1.0.

Sign in with Google.

Connectors configured in a workspace now decide where live evidence is gathered from.

Video citations play in an in-page video player.

Design-partner onboarding arrives with email invitations, member removal and a pilot usage report.

Cursor and OpenCode adapters are runnable, with a command-line tool and MCP registration.

Run detail gains an evidence and citation view, and enterprise administrators get a maintenance-mode toggle with a global banner.

A run that would break a data-residency setting is refused with a clear error instead of starting.

The Codex adapter is runnable.