Review code you don't trust with Open Code Review on a sealed VM
An AI code reviewer in a box you copy, seal, and throw away.
Shelley handles the setup. About 5 minutes, then 2 minutes per review.
Prefer a terminal?
curl -fsSL https://exe-staging.dev/deploy/open-code-review/prompt.txt | ssh exe.dev new --prompt=/dev/stdinOpen Code Review is Alibaba's open-source AI code reviewer (Apache-2.0). It reviews a diff or whole files with an agent that can read and search the repository, and reports comments with a severity, a category such as security or bug, and an exact line. It comes as a CLI, an MCP server, a GitHub Action, and plugins for Claude Code, Codex, Cursor and others. Version 1.13.9 came out on November 13, 2026.
This page uses it for code you don't trust yet: a dependency you're about to add, a contractor's branch, a repository someone sent you. The VM is a review box you set up once. For each review you copy it, clone the code into the copy, cut the copy off from the internet, review, and delete it. The model comes from exe.dev's LLM integration, so no API key is on any of the VMs.
What you get#
- Node.js 24 and Open Code Review 1.13 installed
- Open Code Review using exe.dev's LLM integration, with no model key on the VM
- A connection test that ends with
โ Tool-call round trip verified - No code on the VM; reviews run on copies
What's running, and where
ssh my-open-code-review.exe.xyz
the review box you copy; nothing to review is kept here
http://localhost:5483/ through an SSH tunnel
Open Code Review's session viewer, on a copy
Deploy with Shelley#
Click Deploy and Shelley will set up Open Code Review on a new VM, then check that it works. No config-file wrangling required.
Prefer a terminal?
curl -fsSL \
https://exe-staging.dev/deploy/open-code-review/prompt.txt \
| ssh exe.dev new --prompt=/dev/stdinWhat Shelley is told
Set up Open Code Review on this VM by following the exe.dev recipe at https://exe-staging.dev/deploy/open-code-review.md from top to bottom.
The recipe's example VM name is my-open-code-review. This VM's name is the output of `hostname`: use it wherever the recipe writes my-open-code-review. Skip the step that creates the VM; you are on it.
You are done when all of these are true. Check each one yourself:
- Node.js 24 and Open Code Review 1.13 installed
- Open Code Review using exe.dev's LLM integration, with no model key on the VM
- A connection test that ends with `โ Tool-call round trip verified`
- No code on the VM; reviews run on copies
Run the commands in "Install by hand", from "Install Open Code Review"
on, and the check at its end. Don't clone or review any code on this
VM: it is the template the owner copies for each review. Finish by
telling the owner to follow "Review a repository" from their computer.
Finish with the URLs I should open.
Also at /deploy/open-code-review/prompt.txt. The recipe it follows is /deploy/open-code-review.md.
Install#
On exe.dev, click Deploy to have Shelley set it up on a new VM. Or create it yourself:
Install by hand
On your computer:
ssh exe.dev new --name=my-open-code-review
ssh my-open-code-review.exe.xyz
Install Open Code Review#
On the VM, install Node.js 24 and Open Code Review:
curl -fsSL https://deb.nodesource.com/setup_24.x | sudo -E bash -
sudo apt-get install -y nodejs
sudo npm install -g @alibaba-group/open-code-review
ocr --version
Connect it to the LLM integration#
On the VM, add a provider for exe.dev's LLM integration. The key is a placeholder that Open Code Review insists on; the integration adds the real credentials at exe.dev's edge:
ocr config set provider exe-llm
ocr config set custom_providers.exe-llm.url https://llm.int.exe.xyz
ocr config set custom_providers.exe-llm.protocol anthropic
ocr config set custom_providers.exe-llm.model claude-sonnet-5-5
ocr config set custom_providers.exe-llm.api_key exe-gateway
Check the setup#
On the VM:
ocr llm test
It ends with โ Connection test successful and
โ Tool-call round trip verified.
Keep this VM clean: no code, no GitHub integrations, no tags. Every review starts as a copy of it, and a copy keeps the files, the tags, and any integration attached by tag.
Review a repository#
On your computer, copy the box, clone the code into the copy, then cut the copy off from the internet:
ssh exe.dev cp my-open-code-review review-1
ssh review-1.exe.xyz git clone --depth=1 https://github.com/someone/package.git code
ssh exe.dev egress review-1 integrations
With the egress policy at integrations, the
copy can reach exe.dev's integrations, the LLM one included, and nothing
else. Then review every file:
ssh review-1.exe.xyz 'cd code && ocr scan'
A small library takes under a minute. Each comment names the file and
lines, a category and a severity, then explains the problem and the fix,
and a project summary at the end groups what it found. For instance,
from a review of pallets/itsdangerous:
โโโ src/itsdangerous/url_safe.py:53-53 โโโ
[bug ยท medium] The `serializer` override accepted by this method is silently discarded. This breaks
the documented `Serializer.load_payload(..., serializer=...)` contract for URL-safe serializers
...
For a large repository, start with the parts that matter, such as
ocr scan --path src/auth,src/api, and ocr scan --preview lists the
files a scan would read without calling the model. To keep the results,
write them as JSON and copy them out:
ssh review-1.exe.xyz 'cd code && ocr scan --format json --output ~/review.json'
scp review-1.exe.xyz:review.json ./review-1.json
When you're done, delete the copy:
ssh exe.dev rm review-1
Code on your computer goes in with rsync -a ./package/ review-1.exe.xyz:code/
instead of the clone. Prefer a fresh clone, though: a .git/config from
someone else can run commands when git reads it, and Open Code Review
runs git.
Browse the review#
Open Code Review keeps each review as a session and has a viewer for them, where you can read comments next to the code and mark them fixed or ignored. On your computer, start it on the copy with a tunnel:
ssh -L 5483:localhost:5483 review-1.exe.xyz 'cd code && ocr viewer --open=never'
Then open http://localhost:5483/. Ctrl-C stops it.
Add Codex Security#
Open Code Review is a broad, fast reviewer. For a deeper security pass,
OpenAI's Codex Security writes a threat model
and traces each finding from input to sink; it takes a few minutes and
costs about $0.35 for a small library. To have it on every copy, install
it on the review box, with a profile for the LLM integration. On
my-open-code-review:
sudo npm install -g @openai/codex-security
mkdir -p ~/.codex
cat > ~/.codex/review.config.toml <<'EOF'
model = "gpt-5.6-terra"
model_reasoning_effort = "medium"
model_provider = "exe-llm"
[model_providers.exe-llm]
name = "exe.dev LLM integration"
base_url = "https://llm.int.exe.xyz/v1"
wire_api = "responses"
env_key = "EXE_LLM_KEY"
EOF
Then, on a sealed copy:
ssh review-1.exe.xyz 'cd code && EXE_LLM_KEY=exe-gateway cs scan . --codex "profile='\''review'\''" --headless --max-cost 5'
It prints the number of findings, the cost, and where its report.md
is.
Use it on your own code#
The same CLI reviews your own changes on any VM where you develop; add the provider as in Install. In a repository:
ocr review # staged, unstaged and new files
ocr review --from main --to my-branch # a branch, since it left main
ocr review --commit HEAD # one commit
For every pull request, Alibaba's GitHub Action runs the review in CI and posts the comments inline; its example workflow is ready to copy. With a provider key in your repository's secrets it runs on GitHub's runners. On a self-hosted runner on exe.dev it can use the LLM integration instead, with no model secret in GitHub:
- uses: alibaba/open-code-review@main
with:
llm_url: https://llm.int.exe.xyz
llm_auth_token: exe-gateway
llm_model: claude-sonnet-5-5
llm_use_anthropic: "true"
Open Code Review also has plugins that let Claude Code, Codex and other agents ask for a review; see its plugin guide.
Update#
On my-open-code-review:
sudo npm install -g @alibaba-group/open-code-review
ocr --version
Copies made after that get the new version.
Troubleshooting#
ocr llm testfails:cat ~/.opencodereview/config.jsonshows the provider. The URL must behttps://llm.int.exe.xyzwith protocolanthropicfor Claude models, and the key must not be empty.ocr: command not foundover ssh: it was installed withoutsudo, into a directory only login shells see. Reinstall withsudo npm install -g @alibaba-group/open-code-review.- The copy can't reach the LLM integration after
egress:curl -s https://llm.int.exe.xyz/v1/modelson the copy should answer. On a VM created before exe.dev moved DNS to169.254.169.254, runecho "nameserver 169.254.169.254" | sudo tee /etc/resolv.confonmy-open-code-reviewonce. git clonefails on the copy: run it beforessh exe.dev egress review-1 integrations; afterwards only exe.dev's integrations are reachable.- A review stops partway:
ocr session liston the copy shows it, andocr scan --resume <session-id>continues it.
Frequently asked questions#
How much does a review cost?
The VMs are part of your exe.dev plan, and a copy lives only for a review. The model is billed to your exe.dev account through the LLM integration: reviewing all eight source files of a small Python library with claude-sonnet-5-5 used about 125,000 tokens and took 23 seconds.
Why a copy with no internet?
Open Code Review is an agent: it reads files and searches the repository with tools, and text in the repository can try to steer it. On a copy whose egress is cut to exe.dev's integrations there is nothing to steal and nowhere to send it, and you delete the copy when you're done.
Does it run the code it reviews?
No. It reads the code and asks the model about it; it doesn't build, install, or run anything from the repository. The sealed copy is for the cases where something tries anyway.
Can I use it on my own pull requests?
Yes: ocr review reviews a diff, and Alibaba's GitHub Action posts the comments on each pull request. See Use it on your own code. For a bot that answers commands in pull requests, see PR-Agent.
Can I use a different model?
Any model the LLM integration serves; set it with ocr config set custom_providers.exe-llm.model. Claude models use the anthropic protocol as on this page; for OpenAI models set the URL to https://llm.int.exe.xyz/v1 and the protocol to openai-responses.
Updated Nov 16, 2026 · tested end to end Nov 16, 2026 · Markdown · manifest.json · Something wrong? Tell us