deepsec is an agent-powered vulnerability scanner that runs in your own infrastructure and reviews all the code in an existing repository, rather than only the files changed in a pull request. It targets issues that have been sitting in an application for a long time and were never covered by a review at all.
An init command guides the setup: it asks which model to use, showing benchmark scores and prices to compare, and how to pay for model usage. After that it works unattended, studying the codebase, scanning it, and running the review. The only thing it adds to the repository is a .deepsec/ folder holding its state and findings. An interrupted run resumes from the same command, skipping files it already analyzed, and --max-cost-usd and --max-duration cap what a run may spend or how long it may take.
The pipeline is staged so that model cost falls only where it is needed. scan finds candidate sites with regex matchers and calls no model. process runs the AI investigation and emits findings with a recommendation, and process --diff narrows that to the files changed in a diff for continuous integration gating. triage applies a lightweight P0/P1/P2 classification with a cheaper model, and revalidate re-checks existing findings, including against git history for fixes that have already landed.
It is configured to use models at maximum thinking levels, tunable with --thinking-level, so a scan of a large codebase can cost thousands or tens of thousands of dollars. Model calls route through Vercel AI Gateway by default, which avoids provider-specific keys, or you can pass your own OpenAI, Anthropic, or custom HTTPS provider key with --model-auth direct and no Vercel account. Only the name of the environment variable holding a key is stored, never the key itself. If a run halts because a credential ran out of quota, it stops and reports where to top up.
Large monorepos can fan work out across Vercel Sandbox microVMs with configurable sandbox and concurrency counts. The local working tree is tarballed and uploaded with .git excluded, model credentials stay on the host and are injected only at the selected egress host, and network egress from a worker sandbox is limited to coding agent hosts. The guidance is to treat deepsec like a coding agent with full shell access on whatever machine it runs on, because vendored code and dependencies are a prompt-injection surface.
Features
- Full-repository review: scans all existing code, not only the files a pull request touches
- Resumable runs: re-running the same command continues an interrupted scan and skips files already analyzed
- Cost and time caps: --max-cost-usd and --max-duration bound a single run
- Free pattern pass: scan locates candidate sites with regex matchers before any model is called
- AI investigation: process produces findings with a recommendation, and revalidate re-checks them against git history
- Diff mode: process --diff reviews only changed files, for gating in continuous integration
- Triage: a cheaper model assigns P0, P1, or P2 severity
- Reporting: report, export, and metrics give markdown and JSON summaries, per-finding files, and cross-project counts
- Model routing: Vercel AI Gateway by default, or a direct OpenAI, Anthropic, or custom HTTPS provider key
- Distributed execution: sandbox runs any command across Vercel Sandbox microVMs with credentials held outside the sandbox
- Credential handling: only the environment variable name for a model key is persisted
- Agent-readable docs: the installed package ships skill and documentation files that an agent can read directly
