WORKSPACE GUIDE

A little context.
A lot less guesswork.

The concepts, decisions and everyday habits that make a private AI workspace useful.

Your first project

Start with one clear outcome: a research brief, a document collection or a weekly update. Keep its scope small enough to review.

  1. Name the project and describe its goal.
  2. Add the source material the task actually needs.
  3. Save useful instructions: terminology, audience, format and constraints.
  4. Choose a workflow or describe a task to an agent.
  5. Review the output against the sources before accepting it.

Use separate projects when their source material or permissions differ. Shared workspace references can be selected deliberately rather than assumed.

Files, context and references

PDFs with extractable text, plain text and Markdown are good starting points for document work. Image-only scans need an OCR step before an agent can reliably search their text.

Keep the original source and the prepared text connected. A reference should identify the document and a page or section where possible. If a source changes, refresh its prepared context before starting a new run.

When an answer looks incomplete

  • Check that the intended source set was selected.
  • Check extraction quality, especially tables and scanned documents.
  • Ask for the supporting passage rather than only a confident answer.
  • Keep missing evidence as an open question.

Agent roles and permissions

A role combines instructions, a model and permitted tools. A Reader may only inspect project sources. An Editor can prepare a draft. A workflow can request approval before changing a file or sending information.

Start with read access and draft creation. Add actions only where they are needed. Content inside a source document is task material, not authority to grant new permissions.

Separate instructions from source content and review unexpected requests for new tools, credentials or destinations. A local model can still make mistakes or follow misleading material.

Runs, review and retries

A run moves through visible states so you can distinguish waiting from progress:

  • Queued: waiting for a dependency or available worker.
  • Running: a step is actively preparing context, using a tool or generating output.
  • Needs input: a missing value or decision prevents the next step.
  • Ready for review: output or a proposed action is waiting for your judgment.
  • Completed: required steps and review points are finished.
  • Canceled or failed: future steps have stopped; inspect any existing outputs before retrying.

Cancellation cannot undo an action that has already completed. Review the activity record before rerunning a step that moves files or contacts an external service. Draft work and externally visible actions should have different review requirements.

Models and capacity

The model runs on your workspace host, using the configured local endpoint. Model choice, context length and available memory matter more than the number of named agent roles.

If work is slow, check whether the model is loading, the source is being prepared or the run is waiting for capacity. Reduce simultaneous workers or context size before assuming that more parallel work will help.

The deployment guide covers host planning. Follow your model runtime’s documentation for hardware support, updates and local-only configuration.

Account access and recovery

Use the email address assigned to your workspace account and your password. Your workspace administrator assigns membership and manages account access.

Creating an account

Enter your name, email, password and the invitation code supplied by your workspace administrator. Codes are checked by the server. If a code has expired or is incorrect, ask the administrator for a new one.

Forgotten password or inaccessible account

Contact your workspace administrator. If you are the operator or need help identifying the right contact, email contact@deskarden.com with the workspace address and your account email. Access must be verified before an account is changed. Never include a password or invitation code in your message.

A sign-in error does not confirm whether an account exists. If repeated attempts are limited, wait a minute before trying again.

The operator’s checklist

  • Keep remote browser access behind HTTPS and authentication.
  • Keep the model endpoint private to the workspace host or trusted network.
  • Maintain backups of source files, accepted outputs and configuration.
  • Choose a retention policy for working context, run records and logs.
  • Review tools and outbound connectors before enabling them.
  • Maintain model and runtime licenses, updates and resource limits.
  • Test recovery from a backup, rather than only checking that a backup exists.

A private deployment puts the operator in charge of the environment. See security and data boundaries and our public website privacy notice for their different scopes.

Need a hand? Contact us with your workspace address and a description of what you are trying to do.