Meet Ada, your personal Agent
Ada is the personal Agent inside xopc. Start with a conversation, clarify a goal, or delegate something to do while you keep talking. You can also go directly to the workspace to handle files and projects. Both use the same xopc installation.
Before you begin
Install the desktop app or open the Gateway console, connect a working model, and keep the Gateway running. Voice and external services require separate configuration. Ada is included; you do not need another application or account just for Ada.
Start your conversation
- Open the personal Agent from the sidebar (
#/personal). Its default name is Ada; an existing installation may show a name you chose. - Complete the setup shown on the page and choose an available model if requested.
- Share a little context:
I work independently. Keep replies concise, ask when the goal is unclear, and help me finish one useful thing at a time. - Bring a real question or small task. If you only want to discuss, say so.
The first check is a visible reply in the personal conversation. Confirm that the assistant understood your request before granting access to additional material.
Make it yours
Click the avatar or Response preferences in the page header. You can change the name, appearance, tone, response length, and whether it should listen, untangle a problem, or suggest solutions. Model and voice selection are available in the preferences panel.
Updates controls how much task progress Ada discusses. Proactive messages controls contact about followed topics, including quiet hours. These are separate preferences; use the follow-up guide to choose the latter.
Use the user-understanding controls to review, correct, or delete durable understanding. Changing tone is not the same as correcting a remembered fact.
Delegate when you are ready
Try: Turn our plan into a one-page Markdown file. Include the goal, three next actions, and open questions. Tell me when the file is ready.
Open Activity to inspect delegated tasks. Continue discussing another topic, then review the delivered file and ask for changes. See Task delegation and delivery.
Ada can read short local text, recall earlier conversation details, and inspect notes, projects and automation state. After work returns, you can ask about a particular section, PDF page, spreadsheet cell or source in the saved result without starting the work again. Supported document text is prepared in the background; if a cold read needs preparation, the question continues automatically after it is ready. Visual interpretation, new processing and complete bulk review still use specialist execution. For supported connected apps, Ada checks the connection and account before submitting a read-only request; saved-result reads retain account access checks and do not refresh external data.
In the Web chat, use Ask about beside a returned result or file to attach that exact result to your next message. File previews and downloads remain available while text is being prepared. Extracted text can omit layout, images or spreadsheet recalculation; Ada should keep those limitations visible.
The personal conversation uses a fixed set of local and coordination tools. MCP tools, Codemode and dynamic external discovery run through specialist Agents with their own configuration. Configuring MCP does not give the personal conversation direct access to those tools. A separate ordinary conversation with the same Agent follows its configured tool permissions.
For ordinary delegated work, Ada submits a complete Task brief and returns immediately after local persistence. The Gateway allocates one Specialist in the background. The independent task-router Agent inherits the global chat model unless separately configured; it compares candidate profiles but does not execute business tools. Assignment and execution status return to the originating conversation, and verified results use the existing Task result delivery flow. Explicitly requested Agents and Skills are preserved. Connected-app requests retain their account connection and authorization flow.
The personal model sees submission and existing-task coordination commands. Specialist discovery, matching and direct executor creation are not advertised, and the specialist directory is not injected into its prompt. Describe needed capabilities in the brief; the background router resolves execution conditions. Explicitly named executors can be supplied by ID or unique display name, and explicitly requested Skills are preserved without foreground discovery.
Allocation is durable across Gateway restarts. Before execution, allocation failures can be retried on the same Task with personal_task(command="retry_allocation", taskId=...); answering a router question also resumes allocation. This does not rerun an execution that has already started. The initial router review budget is 24 eligible candidates; larger candidate sets report a clear blocker rather than silently omitting candidates.
If something is missing
| Problem | Next step |
|---|---|
| Personal Agent setup cannot finish | Check the model credential, available model, and Gateway connection |
| The name is different | Review the saved profile; Ada is the default name, not a separate installation |
| A voice call is unavailable | Configure and test Voice; text chat does not establish speech availability |
| An unexpected follow-up arrives | Review Proactive messages and the explanation attached to the topic |
| It has the wrong background | Correct the user understanding and clarify the current conversation |
Core state stays in your local xopc installation. Selected cloud providers process the context sent to them. See Data and file locations and Product philosophy.
Personal tool coordination
Ordinary Tasks support personal_task(cancel) from the main conversation. Cancellation closes the Task and stops further allocation; executionStopConfirmed must be interpreted separately from accepted cancellation. Connected-app Requests use personal_request without selecting an executor in the main model; the Task Router assigns one within the selected account permissions after authorization. Task, Request, Attention and selection queries return paginated, bounded excerpts. See tool optimization and verification.
Task handoffs retain the original request separately from the Agent’s interpretation, pin selected input versions, and allow scoped, paginated reads by the executor. Necessary references to earlier results, notes, Tasks or files can be passed as personal_task.sourceRefs; their authorization and version resolution happen after submission. Background allocation checks required operations, tool policies, registration, model configuration, Skills and account permissions, then checks readiness again before execution. See handoff and capability verification.
Completed ordinary Tasks can now be revised in the same Task, with the original result retained. personal_task(get, section="revision") returns the exact baseline and expected version for atomic revise; execution stays in the background. Result cards show the result revision, and personal_result can still read prior revisions. Final delivery uses an explicit artifact selection, rather than every file published during the run. Saved text can be delivered independently if composition or foreground availability delays it. See revision and final delivery.