Learn Dolmen Desktop

Dashboard

Every agent sits on one rail, grouped under the supervisor that owns it, with its name, role, provider, status and context use shown live.

Attach an agent's structured chat, drop into the real terminal it runs in, or open its Outputs list of the files it created and modified.

A supervisor orchestrates, workers edit code and run builds and notebooks, and researchers browse and write findings into the Library inbox; personas add your own reusable roles.

Agents message one another through the app, every message names its verified sender, and a supervisor can request a reply that is tracked as an obligation.

The Subscriptions view under Activity shows each provider's current allowance readings, such as Claude's 5-hour and 7-day windows and Codex quota.

Plans

Park a tangent as a proposal: a Markdown file that the Plans pane shows in a gallery of open, promoted, parked and closed proposals.

Promote it to a plan and pick a supervisor, which scopes it, sets up a plan folder and has two agents from different providers work out the shape independently, then compare notes, before it packages the work.

Each work package has an outcome, acceptance conditions, owned paths and dependencies, and moves from ready to executing to done.

A worker never completes its own package: the supervisor reviews the report, the Activity turns and the diff before recording complete or needs-work, and code must first land as a commit of exactly the package's declared paths.

The plan tab keeps the plan's documents, a package checklist and each package's attempts, decisions and history, for you and for the next agent.

Activity

The Activity pane lists every agent turn, newest first: who, which task and package, and the paths the app witnessed it touch from observed tool calls, not from the agent's own account.

Filter by agent, path, plan or package, and switch between Time, File, Plan and Flat lenses.

View changes opens a turn's before-and-after diff, and Undo this turn restores its files after a preview, but only while that turn still has both of its snapshots.

Each turn reports its capture health, and changes the app could not attribute to a turn are shown separately.

Decision, closure and lifecycle records are append-only, enforced by the local database, so the record of who signed off what cannot be quietly rewritten.

Files

Edit Markdown in place and leave comments on a selected passage that your agents can read, on PDFs as well.

Word, PDF, CSV, images, code, GeoTIFF, shapefiles and GeoPackage files render inline.

Jupyter notebooks run live on the same kernel for you and your agents, so their runs and yours stay in sync.

The Files, Browser, Plans and Dashboard panes can each be detached into their own window.

Library

Each workspace has a searchable Library of reports, manuals, PDFs, Word and Markdown files, for you and your agents to consult deliberately.

Researcher reports land in an untrusted inbox, and you can add or drop your own files.

Search combines keyword matching with local embeddings from a model that ships with the app, so nothing is sent anywhere to compute them.

Agents cite the passages they use with evidence links that open the exact passage in the Files pane.

Browser

A real embedded Chromium browser shares one tab strip between you and your agents, and each agent tab is labelled with the agent that opened it.

Browser authority is set at launch: researchers can browse fully, supervisors and workers can only open a page for you to look at, and a supervisor can grant a worker a localhost-only toolset for testing its own web app.

Agents can visit only sites on your allowlist, and a site an agent requests does nothing until you approve it.

Login, payment, mail and admin sites are always denied, and clicking and typing need an agent-actions switch that is off by default.

Every agent action and denial is written to an action audit log.