LangStage for VS Code

Chat with your own LangGraph agent from inside your editor. The LangStage panel is the extension's own chat view, so it works in VS Code with or without Copilot, and in Cursor, VSCodium, Windsurf and code-server.

Install: two parts

The extension runs your agent through a small Python process, the sidecar, which comes from PyPI. Installing the extension alone is not enough: you need both parts.

1. Install the sidecar into the Python environment that can import your agent:

   pip install langstage-vscode
   

2. Install this extension from the Extensions view: search for LangStage. Cursor, VSCodium and Windsurf install it from Open VSX.

3. Point langstage.pythonPath at that interpreter if it isn't the python on your PATH, for example /home/me/project/.venv/bin/python or C:\Users\me\project\.venv\Scripts\python.exe.

If the interpreter doesn't have the sidecar, the panel says so and gives you the <python> -m pip install langstage-vscode command for it. The extension expects langstage-vscode 0.5.35 or newer. It still runs an older one, but shows a warning with the upgrade command.

Try the demo

Open the LangStage view in the activity bar. With no agent configured, the panel says No agent configured and offers Try the demo: a keyless demo agent (--demo=tools) for that session, with no API key needed. Include one of these phrases in your message to see each kind of output:

• use a tool shows a tool-call card;
• think shows a reasoning block;
• ask me shows an approval card.

Any other message is echoed back.

Run your agent

Set langstage.agentSpec to your agent: path/to/agent.py:graph or package.module:graph. A relative path resolves against the workspace folder. The agent is any LangGraph CompiledGraph, for example from deepagents:

# my_agent.py  ->  "langstage.agentSpec": "my_agent.py:graph"
from deepagents import create_deep_agent

graph = create_deep_agent(...)

Leave langstage.agentSpec empty to use LANGSTAGE_AGENT_SPEC or the project's langstage.toml ([agent] spec = "my_agent.py:graph"), the same config every LangStage stage reads.

Workspace root

The agent runs in your workspace folder (the first one, in a multi-root workspace) unless you've set a root yourself: a [workspace] root in langstage.toml (a relative path is taken from that file's folder) or a LANGSTAGE_WORKSPACE_ROOT in the environment VS Code was started with. Those win in the same order as everywhere else (langstage.toml, then the environment), so langstage-vscode-sidecar --show-config, run from the workspace folder, shows the root the panel and @langstage use.

Settings

The old deepagent.agentSpec and deepagent.pythonPath settings are still read as a fallback.

Commands

What you get

• Replies stream in as markdown.
• Tool calls appear as collapsible cards with the arguments, the result, the status and the duration. Reasoning appears in a collapsed block. A write_todos plan shows as a live Tasks checklist.
• Approval cards for human-in-the-loop interrupts, with Approve, Reject (with an optional reason), Respond, and Edit (the arguments as JSON).
• Stop cancels a turn, and the agent still remembers the conversation.
• Several conversations per workspace, each with its own memory. Transcripts are saved on this machine and come back after a window reload.
• If you use Copilot, the same agent is also available as the @langstage chat participant, with /approve, /reject, /respond and /edit.

Agent output is untrusted: the panel never renders raw HTML, and a link opens only after you confirm it. The extension is disabled in untrusted (Restricted Mode) workspaces, because running your agent executes workspace code.

Troubleshooting

If the panel says The agent failed to start, it shows the sidecar's startup error, and the full stderr is one click away. To check the interpreter and your agent outside the editor, use the same interpreter as langstage.pythonPath:

python -m langstage_vscode --selfcheck --agent ./my_agent.py:graph

Learn more

• Documentation for the LangStage family
• Full README: configuration, the sidecar's CLI and its stdio protocol
• The sidecar on PyPI
• Report an issue

LangStage runs the same agent on every stage: a web app, JupyterLab, the terminal and VS Code. All of them share the same spec string, langstage.toml and LANGSTAGE_* variables.