Find, confirm and fix security issues in your code without leaving OpenCode.
Ask for a scan in plain words, or drive it yourself from the terminal or the Desktop panel.
The terminal view after a scan. Screenshots use sample data.
It runs Codex Security through the official @openai/codex-security SDK, using the model and provider you already have configured for Codex. This is an early, unofficial project, not an OpenCode or OpenAI product.
1. Add the plugin to the plugins array in ~/.config/opencode/opencode.json:
{ "plugins": [{ "package": "opencode-codex-security@latest" }] }Then restart OpenCode.
2. Open /security in OpenCode. The top of the view shows the provider, model and credentials it will use, and tells you if anything is missing.
3. Ask for a scan, or press n:
Run a security scan of this repo.
You see a preview of what will be scanned and where the source will be sent, and nothing starts until you approve it.
Scans need Python 3.10 or newer.
Have a coding agent install it
Paste this into your coding agent on the machine where you use OpenCode:
Install the OpenCode plugin opencode-codex-security.
1. Add { "package": "opencode-codex-security@latest" } to the "plugins" array of my
OpenCode config (~/.config/opencode/opencode.json, or under $XDG_CONFIG_HOME).
Keep everything else in the file as it is.
2. Tell me to restart OpenCode and open /security to confirm it loaded.
3. If you are running inside OpenCode, after the restart call security_status and
report what it says.
Don't start a scan or change my model provider or credentials. If something is
missing, tell me what it is.
Run it from source
Clone this repository, run npm ci --ignore-scripts && npm run build, and set package to the checkout's absolute path. Run npm run build again after pulling changes.
The plugin gives your session's model the tools and a skill for the whole workflow, so you can describe what you want:
Scan only what I changed on this branch against origin/main.
Scan the billing service. It depends on the auth and ledger repos, and our
Terraform and GitOps repos describe how it is deployed.
Is the tenant check finding real? If it is, propose a fix.
Draft a SECURITY.md for services/api.
For the multi-repository case the model clones what is not checked out locally, reads the infrastructure repositories, writes a short deployment summary, and passes it to the scan as context. When a scan finishes, a notice lands in the conversation and the model reports the findings.
|
Scan Standard or Deep. The whole repository, selected paths, your uncommitted changes, or everything since a branch or commit. For each scan you can pick the reasoning effort, a cost limit and a time limit. Two scans can run at once. Review findings Most severe first, each with its location, evidence and a suggested fix. What you have done about a finding (confirmed, dismissed, fix ready, fix applied) is saved with the scan. Validate Re-check a finding against the source before you spend time on it. Fix The fix is written and verified in a throwaway copy of your last commit. You get a diff, and your files change only if you apply it. Nothing is committed. Draft a security policy Ask for a Discuss Hand any finding to the conversation and have the model explain it. |
There is nothing to configure if Codex already works for you. The plugin uses the model and provider from ~/.codex/config.toml, including gateways and Amazon Bedrock, and authenticates the way the SDK does: an API key from the environment, otherwise your Codex Security sign-in.
To choose a different model, a Codex profile, a default depth or a cost limit, add ~/.config/opencode/codex-security.json:
{
"codex": { "profile": "work" },
"scan": { "mode": "standard", "maxCostUsd": 5 }
}Everything is in docs/CONFIG.md.
/security opens the view, /security-scan starts a scan and /security-validate re-checks a finding. Inside the view everything is a key, and the footer shows the ones that apply to what is selected:
| Key | Does |
|---|---|
n |
New scan, with a review where you set target and scope |
enter |
Move into the findings, then open one |
v / x |
Validate the finding / propose a fix |
y / r |
Apply / discard a proposed fix |
g |
Discuss the finding in the conversation you came from |
e / u |
Run a scan again / resume a stopped scan |
tab |
Switch between scans and other activity |
h |
List every key available right now |
Tools for the session's model
| Tool | Purpose |
|---|---|
security_status |
The resolved route, config files, Codex profiles, and Node/Python found. |
security_scan |
Start a scan. Asks for your approval with the preview. |
security_scans, security_scan_status |
List scans and follow one. |
security_scan_results |
Findings of a completed scan, or one finding in detail. |
security_policy, security_validate |
Start a policy draft or a validation. Each asks first. |
security_fix, security_apply_fix |
Propose a fix for a finding, then apply or discard the diff. Each asks first. |
security_operations |
Status, pending questions and results of those operations. |
security_answer_questions |
Answer a policy's owner questions. Asks first. |
security_cancel_scan, security_cancel_operation |
Cancel one run. |
Approvals use OpenCode's permission rules, under the actions codex_security_scan, codex_security_policy, codex_security_validation, codex_security_fix, codex_security_apply_fix and codex_security_answer.
Context files for a scan may be absolute paths in other checkouts. Any UTF-8 text file (Terraform, YAML, JSON, Markdown), PDF or DOCX works, and a directory contributes all of its text files.
- Your source goes to your model provider. Every scan, validation, fix and policy draft shows where, and waits for your approval.
- Costs are estimates.
scan.maxCostUsdstops a scan on the SDK's estimate; it is not a billing cap. Validation and fixes have no cost limit. - Findings arrive when a scan completes, not while it runs. If a scan fails or is cancelled, whatever it had saved is shown as partial results.
- Standard scans have no time limit unless you set one, and treat it as a target. A scan that lost its connection or was cut off by a restart can be resumed where it left off; any other stopped scan can be run again.
- A fix is written against your last commit. It does not see uncommitted work, and applying it fails cleanly if you have since changed the same lines.
- Change scans run at Standard depth.
- Restarts. Scan results and finding status are kept. A scan that was running is marked interrupted, and policy drafts, validations and unapplied fixes in progress are lost.
- Runtimes. Scans use Python 3.10+ and Node on the
PATHOpenCode runs with. If either is missing or too old, the Security view says so. - Tested on macOS and Linux. The Desktop panel is experimental.
npm ci --ignore-scripts
npm run check # lint, format, types, build, unit tests
npm run setup:local # one-time: isolated OpenCode under .runtime/
npm run test:host # the plugin inside a real OpenCode server, with a mock SDKSee CONTRIBUTING.md, docs/ARCHITECTURE.md and docs/RELEASING.md. Released under Apache 2.0.