Getting started¶
Launch the starter, point it at a JSON-capable model, and walk one synthetic run through pause, restart, approval, and readback.
You need: an AWS account that can subscribe and launch EC2, Session Manager or SSH, and an OpenAI-compatible model that can return a JSON object with exactly two fields. Use synthetic data for the included demonstration.
1. Subscribe¶
Open the LangGraph Workflow Starter listing, choose View purchase options, review the terms, and subscribe.
2. Launch the instance¶
- Choose t3.small (2 GiB of memory) or a larger listed type.
- Use an encrypted gp3 root volume of at least 16 GiB.
- Require IMDSv2.
- Attach an instance profile that can use Session Manager if you connect that way. Leave inbound application ports closed.
3. Connect¶
The API listens on 127.0.0.1:8123. Anyone with the ubuntu account can
control its workflows. This is a single-user service.
4. Configure a model¶
As an administrator, edit /etc/langgraph/provider.json with sudoedit in
a terminal that is not being recorded. Set base_url, model, and
api_key. Include /v1 in the base URL when the provider requires it. Use
HTTPS for a remote provider. A local OpenAI-compatible model may use
loopback HTTP. The provider demonstration requires a remote HTTPS endpoint.
Keep the file root:ubuntu mode 0640, and /etc/langgraph mode 0750.
The service reads the file for each proposal, so a provider change does not
need a restart. Do not put the key in a shell argument or a log.
5. Start, approve, and reread¶
As ubuntu, pick a new result directory:
The tool reads a generated three-row CSV and totals it in code. The model
proposes publication. LangGraph pauses before publication. Read paused.json
in the result directory, and the input.csv and tool-result.json under
the run directory named in that JSON.
sudo systemctl restart sushi-langgraph.service
bash /opt/sushi/docs/demo.sh resume /home/ubuntu/langgraph-results/my-first-run
sudo systemctl restart sushi-langgraph.service
bash /opt/sushi/docs/demo.sh read /home/ubuntu/langgraph-results/my-first-run
Only a successful final read prints SUSHI_CUSTOMER_DEMO_PASSED. It checks
the CSV, the calculator receipt, the report, one calculator call, and one
publication.
For your own bounded integers:
sushi-langgraph start invoice-review-001 5 17 20
sushi-langgraph read invoice-review-001
sushi-langgraph approve invoice-review-001
sushi-langgraph reject invoice-review-001 finishes without publishing.
Reusing an id with different input fails. A rejected workflow needs a new
id to start again.
Last updated November 11, 2026 · Suggest an improvement