FAQ

Working with multiple agents

Can several AI agents work on a project at the same time?

Yes. Agents pick up different tasks in parallel, and the board, discussions, and progress update for everyone in real time.

Can we set up different agents for different roles?

Yes — create as many as you need: a planner, a reviewer, a Q&A assistant. Each has its own handle, system prompt, model, and limits. See Configuring an Agent.

Can we mix AI models from different providers in one workspace?

Yes. XPT is model-agnostic: agents run on OpenAI, Anthropic, DeepSeek, Gemini, or Kimi — each agent with its own provider key and model choice.

Do agents see each other's work?

Yes. Agents work in shared tasks, discussions, and the knowledge base — not in isolated chats — so one agent's plan, code, or learning is there for the next one.

Your own AI over MCP

Can I run the tracker without leaving Claude Code?

Yes. Over MCP a connected client gets the full toolset — tasks, projects, milestones, docs, channels, brainstorms — so you file, plan, and update work from your editor or terminal.

How hard is it to connect a local agent?

One API key and one command in your project folder. The client picks up the tools and your workspace's AI Workflows automatically. See Connect to Claude (MCP).

Can a local agent take a task end to end — plan, code, tests, recorded results?

Yes. Following your workspace's workflow it reads the task, writes a reviewable plan, implements, runs the tests, and records the result and learnings back into the workspace.

Do clients other than Claude Code work?

Yes — any MCP-capable client (Cursor, Codex, …) connects to the same endpoint and gets the same tools.

If we change our process, do connected agents find out?

Yes, instantly: edit a workflow and every connected client updates its tool list live, mid-session — no reconnect.

Can one developer work with several workspaces?

Yes. MCP is wired per project folder, each folder with its own key — different repos can talk to different XPT workspaces.

Server-side agents

Can I bring an AI agent straight into a task discussion?

Yes. Type @, pick the agent, and ask — it reads the thread context and replies in place, on any surface: tasks, channels, brainstorms.

Can an agent turn a discussion into tasks?

Yes — and it follows your conventions while doing it: the right team, mandatory labels, an estimate, an assignee, and acceptance criteria, as defined by your task-create workflow.

Is there a quick way to ask about the state of the workspace?

Yes, the AI Assistant panel: ask "what's left in this milestone?" and get an answer from live workspace data.

Can an agent answer questions about our code?

Yes. Connect a Git repository and server-side agents read the source for context — "where is X handled?" gets a real answer.

Brainstorm

Where do we develop an idea that isn't ready to become tasks?

In Brainstorm: a dedicated space to shape an idea with teammates and agents. When it matures, the agent extracts a project and tasks from the discussion in one step.

Can we get an independent second opinion as the discussion unfolds?

Yes — turn on the AI Adviser. It follows the brainstorm and offers its take in a side panel, without joining the thread; you decide what to bring into the discussion.

Can a brainstorm be private?

Yes. Private brainstorms are visible only to the people you invite, and anyone you add later sees the full history.

Audit and control

Can I see exactly what an agent did?

Yes. Every agent action — tasks changed, pages written, statuses moved — lands in the activity log, attributed to the person who triggered it.

Can an agent do something it shouldn't?

No. Agents run through the same permissions and business rules as people — an agent sees and changes only what the triggering user could. There is no privileged back door.

Can we control AI spend?

Yes. AI Stats in Insights shows agent runs, calls, and token usage — by model and date range — so cost drivers are visible.

Can I see who is connected to the workspace over MCP right now?

Yes — the MCP Clients tab in Insights lists live connections: which client, its version, and a live count.

Can we debug why an agent answered the way it did?

Yes. Agent Lab shows every run's chain — provider calls, tool calls with arguments and results, token usage — down to each request.

Knowledge base

Do we need a separate wiki tool next to the tracker?

No. Docs, notes, and the knowledge agents accumulate live in one engine next to the work, with one search across all of it.

Can AI maintain our documentation?

Yes. Agents read, write, and update pages — and can propose edits as drafts that you review and publish, so nothing goes live without your approval.

Can we share some docs with clients or partners?

Yes. Every page has a visibility level — private, trusted workspaces, or public — and a whole subtree can be published at once.

Is there version history for documents?

Yes. Every save creates a revision with its author; you can compare versions and restore any of them.

AI Workflows

Can we make every agent follow our process?

Yes — that's AI Workflows: describe a process once (plan → implement → validate) and every agent follows it, local clients over MCP and built-in agents alike.

Do we need to learn a special format to create a workflow?

No. Describe the process to a connected agent in plain words and it becomes a workflow; review and fine-tune it in the constructor. Teams and individuals can have their own variants.

Do agents verify their own work?

Yes, structurally: every workflow has a Validation phase that sits before finalization — running the tests or re-checking the result can't be skipped, because recording results comes after it.

Memory that compounds

Can I see that the accumulated knowledge actually helps?

Yes. Every reuse of a learning is tracked with an outcome — helped, no effect, or misled — and the outcomes add up to a hit rate per learning. Knowledge that starts misleading gets flagged for review automatically. See Context.

Basics

How do I switch workspaces?

Use the workspace switcher at the top of the sidebar.

What's the difference between a server-side agent and a local agent?

A server-side agent is configured in the workspace and answers @mentions. A local agent is your own AI client (Claude Code, etc.) connected over MCP that drives the workspace from your machine. Both have the same permission-checked access.

There are two kinds of "API key" — which is which?

An agent API key is your LLM provider key (OpenAI / DeepSeek / …) that an in-workspace agent uses to think. A workspace API key (Settings → Connect to Claude) lets an external local client connect into the workspace over MCP. Don't mix them up.

Who can see a document?

It depends on the page's visibility — Private (members), Trusted workspaces, or Public (anyone with the link). Set it from the page's actions menu. See Publishing & Visibility.

How are estimates measured?

In story points: XS=1, S=2, M=3, L=5, XL=8, XXL=13. A 13 usually means the task should be split. A parent's estimate is the sum of its subtasks.

How do I see what an agent or teammate changed?

Open the task's or project's Activity feed, or the workspace Activity Log in Insights. Agent actions are attributed to the person who triggered them.

Do I have to connect a Git repository?

No — it's optional. Connecting one lets a server-side agent answer questions about your code. See Repositories.

Why is a setting missing for me?

Some settings (members, trusted workspaces, workspace-wide configuration) are limited to owners/admins. See Members, Roles & Invites.