Skip to content

Baselines and the Weekly Digest

This page covers behavior baselines and the Weekly digest in the hosted dashboard, which come with Omnodex Pro. A baseline is what a project’s agents normally do: the hosts they reach, the MCP servers they use and the tools they call. When a session does something the project hasn’t seen before, the dashboard says so, shows how much history that judgment rests on, and asks whether you expected it.

New isn’t the same as bad. Most new things are an agent doing its job a new way. The point is to let you look at the short list of things that changed instead of every call.

Three kinds of item are tracked per project:

ItemCompared across
A host a call reachedThe whole project, on every agent platform
An MCP serverThe whole project, on every agent platform
A toolThe project on one platform, since each agent names its built-in tools differently

An item is new the first time a project’s sessions use it. It becomes normal once it has been used in 3 sessions on at least 2 different days. If a project goes 60 days without using it, it reads as new again.

Calls made by a subagent count toward the session that started it.

A project is the repository a session ran in, read from its git origin remote, so the same repository is one project on every machine, whatever its folder is called. A folder with no remote is its own project on that one machine and is never matched with anything elsewhere. A session the MCP proxy captured outside a repository is grouped by the MCP client that started it.

A project needs some history before “first time” means anything. Until it has 5 sessions and 7 days of history, it is building a baseline: the session shows “Building baseline: 3 of 5 sessions” instead of any new items, and the digest lists the project under “Still building a baseline”.

Open a session to see the New in this project panel, listing what that session used for the first time. Each item says how much history it was compared against, for example “first time in 42 prior sessions over 31 days on 3 machines”, and what was monitoring the agent at the time: “Coverage: Claude Code monitored on this machine since 2026-10-02, no change this week.”

Some items come with an explanation:

  • New coverage, not new behavior. The item showed up within two sessions of that machine starting to see more, such as a hook being installed or the CLI being updated. The agent probably did this before and Omnodex just couldn’t see it. These items don’t ask for an answer.
  • Changed on a server shortly before. A tool called for the first time after the MCP server’s tool list or definitions changed on the same machine says “added on docs 2 days before”.
  • Right after a model change. A first sighting in the sessions after the model changed on that machine names the new model.

The Risk Events panel also notes when a finding’s call touched something new in the project.

For each new item, answer Did you expect this?

AnswerWhat happens
ExpectedThe item is normal for this project from now on
Not expectedThe item is held: it keeps showing as new, in the session and in the digest, until you change your answer to Expected
No answerIt stays new until it has been used enough to become normal

An item with a HIGH or CRITICAL finding on it is held too, however often it’s used, until you mark it Expected. Answers are kept in your account’s encrypted policy, so they apply on every machine and from any browser.

Weekly digest in the sidebar is one week, Monday to Sunday, in the time zone you chose in Settings. Previous week and Next week move through the last 120 days. The current week says “in progress”. The top line counts the week’s sessions, machines and projects, and names any project still building a baseline.

SectionWhat’s in it
Needs reviewHeld items, new items nobody has answered yet, and HIGH or CRITICAL investigations with no action taken (not muted, and no policy entry for the rule). Answer Expected or Not expected here, or open the session
What changedEvery new item, server change and change to monitoring in the week, newest first. New items carry their sample, and the ones explained by coverage say New coverage. The first 50 are shown, with Show all for the rest
Recurring chainsA rule and destination that turned up in two or more sessions this week, with how many sessions had it last week
Noisy findingsUp to five rules with at least three findings this week, with how many per 100 sessions. Expand one to act on it with the same finding actions as the rest of the dashboard
CoverageWhen the week’s monitoring was less than complete: a platform not fully monitored, a machine not reporting with a session open, and recording or upload pauses

The digest exists so you don’t have to read every finding. It answers three questions: what needs a decision from you, what changed, and how much of the week Omnodex could actually see.

  1. Start with Needs review. These are the only items waiting on you. Answer Expected or Not expected for each new item, or open the session when you can’t tell from the row. When the section says Nothing is waiting on you this week, you’re done with the part that matters most.
  2. Skim What changed. New hosts, servers and tools, changes to MCP servers, and changes to monitoring. Anything marked New coverage is new because Omnodex started seeing more, not because an agent did something new.
  3. Look at Recurring chains. A chain that shows up in several sessions is a habit, not an accident. If it’s expected, allow it from one of its findings so it stops coming back.
  4. Tune Noisy findings. A rule firing on most sessions usually needs an allow or mute scoped to a path or project, which you can add from the row. Each one goes to your Policies.
  5. Check Coverage last. It tells you how far to trust the rest. A machine that stopped reporting with a session open, or a long pause, means the week may hold more than the digest shows.

The digest is worked out each time you open it, so a past week shows whatever history is still in your data. Nothing in it calls a change malicious: it says what was observed and how much history backs it.

Baselines are worked out from the data your machines sync, not from the live view. With automatic sync, that means:

  • Claude Code, Codex, Antigravity and other hook platforms: when the session ends and its data syncs.
  • Cowork Desktop, ChatGPT Desktop and other MCP proxy sessions: every 15 minutes while the session runs.

The dashboard draws the session first and adds the baseline once it’s ready, so the panel can appear a moment after the rest of the page.

The baseline and the digest are worked out in your browser, after it decrypts your data. The result for each upload is encrypted with your stream’s key and stored beside that upload, so the next load can reuse it, and it’s deleted along with the data it came from. The hosted service learns only that a stored result exists, its size and which analysis made it. Your answers are part of your encrypted policy. See Data Handling.

The local dashboard doesn’t have baselines or the digest.