Agent Graph

Agent Graph is a multi-agent engineering workspace for Visual Studio Code and Code Server. Describe a goal to the Manager Agent, turn it into focused tickets, and watch specialized agents plan, build, review, test, and deliver the work through a visible graph.

Start here

Open Agent Graph from the Activity Bar and select System Chat, the first row under Projects. Then ask:

How do I use Agent Graph?

The Manager Agent can help connect your model provider, explain the current team, tune the graph, add a project, and dispatch your first task.

Highlights

• One durable Manager Agent — VS Code and Code Server share the same background session across all your projects.
• Visible multi-agent work — follow live thinking, tool activity, handoffs, results, and errors without leaving the editor.
• Parallel tickets — run independent tasks at the same time while Agent Graph coordinates shared agents and final reviews.
• Dynamic agent graphs — workers can route work to the best next specialist while every required agent still gets involved.
• Codex and Pi executors — choose the container runtime, model, and reasoning depth for each agent.
• Safe project isolation — Git-backed tickets work in separate worktrees; approved changes are reviewed before integration.
• Durable background execution — active work continues through editor reloads, browser refreshes, and closed windows.
• Built-in or self-hosted inference — use an isolated Codex subscription or connect an OpenAI-compatible vLLM Responses endpoint.
• Stable MCP control plane — the built-in Manager and trusted clients use the same documented tools through one persistent, authenticated daemon.

How it works

The Manager Agent turns your request into one or more focused tickets and selects the specialists each ticket needs. Agent Graph schedules the work, shows live progress in the chatroom, and keeps the complete ticket history in the Issues view.

Several tickets can progress together. Each ticket keeps its own workspace and agent discussion, while Agent Graph prevents the same agent identity from being used by two tickets at once. When integration is required, the selected final reviewer checks the result before Agent Graph merges it back into the project.

The sidebar keeps the workflow simple:

• Projects — open System Chat, project chatrooms, and issue lists.
• Agents — review the team, models, reasoning depth, and executor choices.
• Graph — see routes and live agent state.

Setup

1. Open Agent Graph → Settings.
2. Keep Codex subscription and complete Login, or select vLLM and enter its URL and API key.
3. Load the model list and choose the Manager Agent model and depth.
4. Configure Docker and build the Codex and Pi worker images.
5. Return to System Chat and ask the Manager Agent to review or optimize your team.
6. Add a project and describe the work in its chatroom.

Docker Bridge mode is recommended. Use Host only when your host can reach a vLLM service that Docker bridge networking cannot, such as on some VPN or policy-routed networks.

Chatrooms and issues

Chatrooms keep the human-facing conversation clear. Manager thinking and tool calls appear in the status area above the input, while active graph workers appear in the overlay below the header. Useful final agent responses are added to the conversation with expandable details.

The Issues view is the durable ticket record. Open a ticket to see its description, graph queue, current activity, agent reports, comments, and final outcome.

You can drop files onto the chat input. Agent Graph stores the attachment under its private root and adds a safe file reference to the Manager request.

Privacy and isolation

Graph agents run inside managed Docker containers. A worker receives only the workspace and issue context required for its ticket, plus that agent's private notes. Agent Graph settings and credentials stay outside worker containers.

Subscription login is isolated from your normal Codex configuration. Agent Graph does not read or modify ~/.codex.

Graph workers have write access to the ticket workspace mounted inside their container, so use trusted projects and prompts.

Local MCP access

Trusted local agents can operate Agent Graph through the same MCP tools used by the Manager Agent. Connection details are stored in ~/.agent-graph/status.json. The owner-only MCP capability is retained across ordinary daemon restarts, editor reloads, and extension upgrades, so configured clients do not need a new token after every restart.

The loopback MCP service can also sit behind an authenticated same-host HTTPS reverse proxy. Custom proxy hostnames are accepted after loopback and bearer authentication; browser origins remain restricted for safety.

See the MCP operator manual for practical examples covering agent graph updates, parallel tickets, monitoring, cancellation, and recovery. MCP clients can also call the read-only tutorial tool to load the manual, or use ask_manager, cancel_manager, and reset_manager to operate the shared Manager session.

Requirements

• Visual Studio Code or Code Server 1.127 or newer
• Node.js 22.19 or newer in the extension host
• Docker for graph-agent execution

Code Server browser sessions must use HTTPS or a localhost address. Browsers do not enable the service workers required by VS Code webviews on plain-HTTP LAN addresses such as http://192.168.x.x, so Agent Graph editor views will remain blank on those origins.

Settings

• agentGraph.layout selects a flow or radial graph layout.
• agentGraph.showInactiveAgents shows or hides inactive agents.

Agent Graph 0.3.4

Version 0.3.4 makes MCP integrations easier to operate. Agent Graph now keeps one owner-protected MCP capability across normal restarts and upgrades, supports authenticated same-host reverse-proxy domains, and preserves the durable Manager, parallel ticket scheduler, and isolated Codex/Pi execution model.

Feedback and issue reports are welcome through the project repository.