LangGraph Workflow Starter
LangGraph 1.2.18 on Ubuntu 24.04 runs one stateful workflow: a tool totals a CSV, a model proposes publishing the result, and the workflow pauses until you approve. Checkpoints are in SQLite. There is no browser UI. LangSmith Server, Studio, and commercial control-plane features are not included.
- AMI release
- 0.1.1 (LangGraph 1.2.18)
- Recommended
- t3.small
- Storage
- 16 GiB+ encrypted gp3
- Access
- Session Manager or SSH
- Released
- Oct 30, 2026
- Delivery
- Amazon Machine Image in your own AWS account
LangGraph OSS with local SQLite checkpoints and a bounded CSV calculator workflow.
Human approval before report publication, with checkpoint resume across graceful service restarts.
Authenticated loopback JSON API, customer CLI and an independent CSV/result readback demonstration.
Quick start
Full guide →What is on the image¶
This guide matches Marketplace package 0.1.1 (LangGraph 1.2.18). Check the listing for the version currently on offer.
- LangGraph OSS 1.2.18 and a local JSON API on
127.0.0.1:8123. - A bearer token in
/home/ubuntu/.config/sushi/langgraph/access.json, mode0600. The client does not print it. - SQLite checkpoints and run files under
/home/ubuntu/langgraph-data, mode0700. - Demo readbacks under
/home/ubuntu/langgraph-results, mode0700. - Provider settings in
/etc/langgraph/provider.json. - The workflow in
/opt/langgraph/workflow.py.
The supplied tool can read its run's CSV and publish its report. It cannot run a shell command or open a path you choose. LangSmith tracing is disabled.
Recommended instance¶
t3.small (2 GiB of memory) or a larger listed type, with an encrypted gp3
root volume of at least 16 GiB and IMDSv2. No GPU and no model weights are
included. The listed types are t3.small, t3.medium, t3.large,
t3.xlarge, and t3.2xlarge. Leave inbound application ports closed.
Forward port 8123 only if your own client runs on a workstation. A browser
opened on that port does not show a UI.
Scope and limits¶
The model must support OpenAI-compatible JSON-object responses and must
return exactly action set to publish_report and run_id equal to the
workflow id. Markdown fences, extra keys, and non-finite numbers are
refused. Integers are limited to an absolute value of 1,000,000, and a
workflow holds at most 100 rows. Ids are 8 to 64 ASCII letters, digits, or
hyphens. The API serializes requests and supports one local process. The
restart checks in the demo are graceful service restarts, not a claim of
exactly-once execution across power loss.
The starter sends the calculated sum, the row count, the run id, and an input digest to your provider. It does not send the original CSV rows.
Billing¶
The listing states USD 0.03 per running instance-hour. Confirm the offer, including its free trial, before you subscribe. EC2, EBS, data transfer, and model-provider usage are billed separately.
Where to go next¶
- Getting started: configure a model, pause, approve, and reread.
- Configuration: the provider file and the HTTP routes.
- Operations: backups, token rotation, and support.
- Troubleshooting: 401, 409, and 503.
Upstream project¶
LangGraph OSS is MIT-licensed. Software Sushi does not claim upstream
endorsement. Sources and wheels are under /opt/langgraph/sources and
/opt/langgraph/wheels. Notices are under /usr/share/doc/langgraph. See
the LangGraph documentation
and the public customer guide.
Last updated November 11, 2026 · Suggest an improvement