Multi-agent security audits of pinned source code. Agents investigate and test; the supervisor schedules work, validates outputs and saves evidence. Audits run in isolated Cloudflare containers with an EVM toolchain.
Run an audit · Architecture · Operations · Contributing
| Stage | Purpose |
|---|---|
| Preflight | Check source, tools and baseline tests without a model. |
| Discovery | Investigate possible vulnerabilities over research waves. |
| Synthesis | Group observations without losing their evidence. |
| Coverage | Independently look for missed issues. |
| Negative review | Check observations and the reasons they were kept or dismissed. |
| Verification | Test eligible observations against their original conditions and evidence. |
| Reporting | Present supported findings, refutations and unresolved questions. |
Agents use Codex, DSH or Claude SDK runners to inspect source, execute tests and search the web through Exa MCP. The image built from the Dockerfile includes Foundry for building and testing Solidity. Model pools assign runners and models; search settings control which decisions require search. Returned evidence helps agents test concerns against target code. Each task starts a fresh conversation with source and prior evidence as files; output corrections continue in that conversation.
Accepted output has passed format and evidence checks; that does not establish that its security claims are correct. Reviewers do not vote away earlier evidence: references kept at any earlier stage, and conflicting review outcomes, go to Verification. Grouping related observations preserves their individual evidence.
Completion means the selected stages finished; a discovery-only run has no reviewed report. Even a reviewed report can contain unresolved questions. Architecture explains these decisions.
Use a provisioned Swarm service; for a new install, follow First-time setup first. Copy the example configuration and prompt file to your private folder (template), then set your source, models and budgets. Choose stages independently of research prompts:
{
"method": "security-audit-v1",
"stages": ["discovery", "synthesis", "negative-review", "verification", "reporting"],
"promptFile": "prompts.json",
"forceSearch": true,
"forceSearchStages": ["discovery", "negative-review"],
"limits": { "maxDiscoveryRuntimeMs": 18000000, "maxModelCalls": 100, "maxCostUsd": 20 }
}Publish source separately from the container image. Discovery limits stop new research; admitted work and later stages still finish. Follow Run an audit for credentials, source upload, publication and start/stop commands.
| Cloudflare service | Purpose |
|---|---|
| Workers | Serve the API and dashboard and execute audit runtime code. |
| Workflows | Coordinate long-running audits across execution steps. |
| Durable Objects | Coordinate release activation and runtime reservations. |
| Containers and Container Registry | Run isolated agent workspaces from pinned images. |
| D1 | Index audits, releases and runtime assignments. |
| R2 | Store source, releases, evidence and checkpoints; keep Terraform state separately. |
| Secrets Store | Supply credentials without embedding them in source. |
| Access | Protect dashboard sign-in. |
| Workers Logs and Analytics | Diagnose failures and inspect resource usage. |
Architecture shows how these fit together. Code deployment uses Wrangler/CLI/API; infrastructure resources and new services belong in Terraform. The included Terraform configuration defines the resources; Operations covers deployment and image builds.
Install the Node.js and pnpm versions specified in package.json and python3, then:
pnpm install --frozen-lockfile
pnpm checkThis runs type checking, tests and a Worker dry build, without paid model calls or
container builds. pnpm audit0:acceptance runs only the deterministic audit-flow tests.
There is no supported end-to-end local audit server: the engine's dev command
uses wrangler dev --remote. See Contributing for code navigation
and change validation.
Apache-2.0. See Security for reporting vulnerabilities and handling private source, credentials and audit results.