Start from Connect, pick your source host, and Sebastion runs three scanner classes on every PR/MR (AI code audit, regex-based secret detection and OSV.dev SCA). Findings land inline with severity and CWE id, each with a ready-to-apply fix. Free for public repos.
Install once, pick a repo, and we backfill your most recent PR so you can verify output quickly. Future PRs are audited on open / push. Plan limits stay transparent in pricing and rate limits.
Cloning repository at PR head…
Sebastion is an agent. It composes deterministic scanners (OSV-Scanner and regex secret detection) with frontier-model reasoning to turn diffs into actionable security findings.
Sebastion knows your stack (Django, FastAPI, Flask, Next, Express, Go, Spring, Rust) and applies the framework-specific vulnerability checklist on every PR. Catches injection, broken auth, insecure crypto, unsafe deserialisation, SSRF, path traversal and the long tail of OWASP Top 10 issues with the diff as context.
Regex-based detection for AWS access keys, GitHub PATs (classic + fine-grained), OpenAI / Anthropic / Stripe / Slack / Google API keys, JWTs (with header validation), and private-key blocks. Diff-additions only, with EXAMPLE / placeholder filtering to keep the false-positive rate low.
When `package.json`, `requirements.txt`, `go.mod`, `Cargo.toml`, `Gemfile.lock` or any supported manifest changes, Sebastion queries OSV.dev for known CVEs against the new versions. Real CVE ids, NVD advisory links and ready-to-apply fix versions land as findings on the same PR.
Findings post as line-level review comments on the offending change, tagged with severity emoji and a MITRE CWE link. Critical findings request changes on the PR so it is blocked from auto-merge. No new dashboard to learn, no email digest, no separate scanner repo.
Drop a .sebastionai.yml at your repo root to gate severity, ignore paths, suppress rules or disable whole scanner classes. Defaults are sensible. A noisy monorepo never has to uninstall.
severity_threshold: high ignore_paths: - "**/*.test.ts" - "vendor/**" disable_rules: - missing-helmet disable_scanners: [] # llm | osv | secrets review_event: on_critical: request_changes on_high: comment default: comment
Same pipeline on every push. Re-runs incrementally when you push a fix, so the inline comments stay in sync with HEAD.
Open app.foundationmachines.ai/connect and choose your source host. Connect supports GitHub and GitLab SaaS.
On every PR event, Sebastion fetches the diff via the GitHub API and runs the full review pipeline. No source clone, no persistent state. The diff stays in memory for the audit run only.
A walkthrough comment groups the changes by theme and estimates review effort. A pre-merge checks panel verifies the PR title is descriptive, the description is complete and any linked issue is actually addressed by the diff.
AI code audit for SAST in parallel with OSV.dev for dependency CVEs and regex-based secret detection over diff additions (around 20 high-confidence detectors covering AWS, GitHub PATs, OpenAI, Anthropic, Stripe, Slack, Google API keys, JWTs and SSH keys).
Findings post on the offending lines with severity, CWE link, prose explanation and a suggested fix. One-click GitHub suggestion blocks where the patch is small. A copy-paste prompt block on every comment so you can hand the fix to Claude or Cursor. Critical findings request changes so the PR cannot auto-merge.
Each new push re-audits only the changed commits. The first audit on a repo silently baselines existing findings so a freshly installed Sebastion never avalanches an old codebase with hundreds of legacy comments. Only new debt introduced by future PRs gets flagged.
Reply to any inline finding with @sebastionai and the bot responds in the same thread — it answers your question, suppresses the rule on that file, or both. The next audit honours what you taught it.
@sebastionai help @sebastionai ignore # suppress this rule on this file @sebastionai suppress legacy code, rewriting # …with an optional reason @sebastionai mute training data not real @sebastionai why is this exploitable? # Pro+ Q&A — answers in-thread
Ask why a finding flagged, what the fix is, or whether it applies to your context. The model gets the finding + the snippet around it, so the answer is grounded in your code — not a generic explainer.
Reply @sebastionai ignore on a false positive and that (rule, file) is silenced for your installation. Strict severity floor: chat can suppress medium/low/info only — critical and high still require an operator override with an alert.
Sebastion auto-reads AGENTS.md, CLAUDE.md, .cursorrules and .github/copilot-instructions.md from your base branch. Add # sebastion: ignore <rule> in <glob> and the bot honours it on the next audit.
Each tier wires Sebastion to the right model for the job. Provider-routed so we can swap the underlying model vendor without changing your install.
Public repos, framework-aware AI, OSV CVEs and secret scan.
Fast frontier model for SAST.
Open ConnectPrivate repos, higher-reasoning model, autofix PR drafts.
Higher-reasoning frontier model for one developer.
Open ConnectUnlimited users and repos. Top up credits when you need them.
Frontier model with shared credit balance + spend cap.
Open ConnectFull feature comparison and credit calculator on the pricing page.
Sebastion installs in a click and reviews every PR from a contributor. No new build step, no separate scanner repo, no badge fatigue. Public repos are free forever.
AI codebases ship fast and the model writes a lot of the diff. Sebastion is the second pair of eyes that doesn't get tired. It calls out the SQL injection, the leaked key and the vulnerable dep before anyone hits squash and merge.
Team covers every repo and every developer under one shared credit balance — no recurring fee. Top up in $25 / $100 / $500 packs (with bonus credits on the larger ones), turn on auto-recharge if you don't want to think about it, and set a spend cap so PR volume scales without per-seat admin.
Sebastion never persists your source. PR diffs are fetched via the GitHub API, routed through scanners and discarded as soon as the audit completes. We never train models on your data and no-train flags are set on every upstream request that supports them.
Open Connect, pick your source host, and review your next pull or merge request in under a minute. Free for every public repo.