Catching issues in your database infrastructure before they turn into data loss, service downtime, a breach in security, or similar much bigger issue is difficult to stay on top of. It often takes a seasoned database engineer (or a team of them) to be able to achieve a complete picture of the state of your infrastructure, especially at scale, while catching deviations that help you prevent real problems before they occur.

From our PostgreSQL engineers to your infrastructure, our open-source, agentless, self-hosted tool, the pgEdge AI DBA Workbench, is designed to help you get real, in-depth insights across your database landscape. Our team just released 2.0.0-beta1, which - amongst other feature additions, security enhancements & more - introduces smarter anomaly detection and alerting along with support for optional decision models like TypeSafe’s Jev, which make the yes/no call on an alert faster at a fraction of the cost of a reasoning LLM with an explanation as to why.

Stay Aware… Without the Noise

Threshold alerts in the Workbench now have hysteresis: by default, a metric must cross its limit on two consecutive samples to raise an alert, and stay inside it for three to clear. This means alerts won’t continuously trigger, no longer flapping on and off around the limit. Several built-in rules have been re-tuned to prevent false alarms, and anomaly baselines now cover more.

When an anomaly alert shows up, a “Why was this flagged?” button explains the verdict, and a baseline comparison chart plots the metric against its normal range so you can see how far it drifted from usual behavior.

The Workbench supports a wide variety of LLM models for covering your needs with a third-party LLM - such as Claude, OpenAI, Gemini - or one you locally host yourself like Ollama, LM Studio, llama.cpp, EXO, or others. In this release, you can now use a decision model like Jev or OpenAI’s GPT 6 Luna decision model variant. Or, you can turn AI off completely. It’s your choice.

The Workbench has built-in checks for both. It raises an alert when a probe that had been collecting stops, or falls more than three collection intervals behind, and it raises a separate system alert when an AI provider keeps failing.

Alerts Can Be Faster, AND Cheaper

Our team decided to introduce decision model support as the use of such models come with distinct advantages - namely, when it comes to speed and costs.

The Workbench already sets out to be efficient: statistical baselines establish normal behavior per metric, per instance. Embeddings are created that represent the behavior of the system on an ongoing basis; when a value deviates from the norm, it raises an alert. During the final call, anything that passes through tier 1 and tier 2 is evaluated by the LLM/DM; they are much more capable and can look at a lot more context, but are much slower and more costly than the lower tiers. The intent is to filter out as much noise as possible before the alert ever reaches this stage. If AI is turned off or if the model is unavailable, the alert will go off by default.

In this final step, if you are using a reasoning LLM such as Claude, you’ll receive a paragraph of analysis. A decision model will answer with a probability - a much more concise bit of output! In that case the LLM is called only when the alert detail is accessed for the first time by a human, so you don’t end up paying for the API call for alerts that resolve themselves before you look at them or ones you don’t look at all.

We put it to the test against Claude Haiku 4.5, where (as of October 2026) Jev agreed to Haiku’s verdicts 89% of the time. It took about 200 ms per decision instead of ~ 3 s, with a cost of $0.04 per 1,000 decisions instead of $1.93, and never suppressed an alert that Haiku let through.

If the decision model fails or times out, the anomaly goes to your reasoning model - if that fails too, the alert is raised anyway. Decision models are optional. It should be noted Jev is a hosted service that needs a TypeSafe API key, so anomaly context leaves your network when you have this enabled.

But Wait, There’s More!

A slew of other features went into the latest release.

On the topic of integrations: notifications can now be sent out via Telegram alongside email, Slack, Mattermost, and webhooks as designated channels. As a bonus addition, we’re also excited to say there is now pgEdge ColdFront support added thanks to Paul Rothrock, meaning metrics can now be stored in Apache Iceberg tables on object storage for long-term retention at lower cost.

Richer dashboards are now available, with easily accessible views like Top Queries on the Database Dashboard, plus sorting, filtering by query text and database, and pagination on both dashboards. There’s also more detail when sorting through query results, including minimum, maximum, and average timings along with the client that last ran it. New server dashboard sections have been added like Connections, which can now be grouped by user, client, or database, broken down by state, and Replication Slots. Dashboards and the event timeline let you line graphs up with a known incident with custom time ranges.

Ellie, the AI assistant also saw some important upgrades; SQL that Ellie suggests can be run with one click against the right server, and there is more control over LLM requests through allowing reasoning effort, prompt caching, and extra request options to be set. Default models have been updated.

No release would be perfect without a slew of enterprise and security patches. In this release, one of the top changes is that single sign-on is now available through OpenID Connect. There is also now a tamper-evident audit log of permission changes: every change to users, tokens, groups, and permissions is recorded with a hash chain that shows if the log has been altered. Passwords and secrets are masked in more scenarios for stronger credential protection, API token scopes and connection visibility have been hardened throughout, and there is safer AI access to monitoring data as well as safer package upgrades.

Before You Upgrade

2.0.0-beta1 is a real upgrade: PostgreSQL 15 or later is required, schema migrations only go forward, and there are breaking changes such as a security-related change to how the query_datastore MCP tool takes arguments. So take your time: try out the upgrade on a non-production datastore first and take backups before you start. We have an upgrade guide available - take a look.

For the full list of changes in this release with relevant details, view the release page; if you're ready to get started, setup instructions can be found in the docs.

A special thanks to Dave Page who created the Workbench and our engineers who additionally contributed to this release. We’re curious to get your thoughts - review the code and leave issues or pull requests if you’d like to contribute, or get in touch with us directly via Discord or email, anytime.