Agent Dev overview

Agent Dev is the agent that works on your machine — in your repository, in your terminal, and next to your Axelera hardware. Here is the surface in one look, and the path through it.

Agent Dev is the on-device agent inside the Axelera Wingman desktop app for Linux. Claude Code, Codex, or Kimi runs locally on your machine, next to your repository, your terminal, your Voyager SDK venv, and your Axelera cards. It sees what your terminal sees: workspaces, axdevice, /dev/metis*, and the SDK environment. Hosted Chat plans and reviews; Agent Dev builds and runs.

Chat panel and dashboard

Two panels do the work. The Agent Dev chat panel is where you talk to the agent: write a request, watch it run, approve commands, and review changes. Beside it sits the Axelera Dashboard — the same web dashboard you can open in a browser, embedded in the desktop shell. Inside the app the dashboard gains local context: Agent Dev owns the terminals and workspace files, while the dashboard supplies knowledge, projects, MCP setup, and settings.

Welcome page and navigation

The dashboard opens on Home with shortcuts for Wingman Webchat, Knowledge Base, Port to Axelera, and Projects — when Projects is enabled — plus Connect Wingman MCP and Settings. The cards fill the available dashboard area and rearrange into more rows in a tall pane or fewer rows in a wide, short pane. In tight spaces they show icons with their names still visible underneath; larger panes also show descriptions. The visible Get Started button opens the setup guide in either layout. The Axelera and Wingman logos, connection indicator, and account icon stay in the top bar, with no large introduction above the shortcuts. A dock along the bottom keeps Home, Knowledge Base, Port to Axelera, Projects, and Settings reachable from other pages.

Your path through Agent Dev

Your first session: Set up a provider, check the compact readiness dot, and send your first prompt. If it expands to Fix now, click for the guided fix.
Providers, models & effort: Choose Claude, Codex, or Kimi, pick a model, and set how hard it thinks with reasoning effort.
Run & review: The daily loop: attach context, approve commands and patches, and read diffs before they land.
Build your Knowledge Base: Crawl docs, upload files, and index repositories so the agent answers from your sources.
Projects: Collect what an agent writes for one effort, and browse it as folders and files.
Desktop shell & feedback: Learn the chrome around the panels — and how to tell us when something breaks.

Requested and running model identity

The model label starts as Requested followed by the exact model ID you chose. When the active CLI reports its observed identity, the label changes to Running followed by that exact ID. If the CLI has started but supplies no identity, the label says Running model not reported. Wingman never presents a requested identity as observed fact.