Installation¶
Each LangStage stage is its own PyPI package, so install only the surfaces you
want. They all share langstage-core, which is pulled in
automatically.
Or search LangStage in the Extensions view. VS Code installs it from the
VS Code Marketplace; Cursor, VSCodium and Windsurf install it from
Open VSX. The .vsix on the
latest extension release
is the fallback (Extensions: Install from VSIX…). Its LangStage panel works in
VS Code without Copilot, Cursor, VSCodium, Windsurf and code-server; see
VS Code.
Requirements¶
- Python 3.11 or newer, for every package including the core.
- Your agent's own dependencies. Every surface already depends on
langgraph, so a plain LangGraph graph (and every--demo) works on a bare install. If your agent usesdeepagentsor a provider package such aslangchain-anthropic, install those yourself. - An API key for whatever model your agent calls (for example
ANTHROPIC_API_KEY). Not needed for the demos or the keyless preflights.
langstage run with no agent
With no --agent and no configured spec, langstage run falls back to a
built-in default agent. That agent needs pip install "langstage[deepagents]"
and ANTHROPIC_API_KEY. Use --demo for a zero-setup first run instead.
Check it works, no key needed¶
langstage run --demo # web UI at http://localhost:8050
langstage-cli --demo "hello" # one-shot terminal reply
Both run the built-in echo agent (langstage_core.demo.stub:graph), so a fresh
install is provably working before you wire up a real agent. Every surface has a
keyless demo:
| Surface | Keyless demo |
|---|---|
| Web | langstage run --demo |
| Terminal | langstage-cli --demo "hello", or langstage-cli -a langstage_core.demo.tools:graph "use a tool" for tool calls, reasoning and an approval prompt |
| JupyterLab | langstage-jupyter --demo |
| VS Code sidecar | langstage-vscode-sidecar --demo --message "hello", or --demo=tools |
| Reference agent | langstage-hermes demo (a subcommand, not a flag) |
| AG-UI server | langstage-agui --demo, or --demo=tools |
Preflight your own agent¶
Once you point a stage at your agent, each surface has a preflight that loads it and runs one real turn, so a typo'd spec, a missing key or a broken tool shows up before the first chat:
| Surface | Preflight |
|---|---|
| Web | langstage check --live (drop --live for the fast, keyless static checks) |
| Terminal | langstage-cli --verify |
| JupyterLab | langstage-jupyter --verify (the agent) or --serve-check (the HTTP endpoint) |
| VS Code sidecar | langstage-vscode-sidecar --selfcheck |
| Reference agent | langstage-hermes verify (and doctor for the environment) |
| AG-UI server | langstage-agui --verify |
They all exit 0 when the agent is healthy and non-zero when it isn't. See
Exit codes.
Next: Quickstart →