Article

When to self-host Vigilo for AI security reviews

For software and security teams, assess Vigilo’s model access, hosting costs, licence terms and alpha support before planning a controlled trial.

Editorial illustration for When to self-host Vigilo for AI security reviews: a document and lock represent artifact protection. Not documentary evidence.

CERT-EU announced the public release of Vigilo on 13 November 2026, giving software and security teams an open-source system that coordinates AI agents to review applications. A controlled trial makes sense if your team can operate it, approve where model processing happens and validate its findings. Self-hosting the interface does not automatically keep that processing on your servers.

The first constraint is maturity: Vigilo’s security policy calls it alpha and marks versions below 1.0 as unsupported. Treat it as an additional review tool to evaluate, not a supported production dependency or proof that an application is secure.

Do you have source code, or only a running application?

The repository documentation describes two routes. Choose the route before choosing an interface:

  • With source code: white-box review accepts a repository and, optionally, a running copy of the application. It can propose patches. Both the web platform and command line offer this mode. Without a live target, checks stay at source level; that does not establish whether a flaw is reachable in a deployed service.

  • Without source code: grey-box review takes a live application URL and optional login credentials. It produces a report and detection hints, not source-code fixes. This route is currently command-line only.

Live testing is active: the documented workflow attempts exploitation, rather than merely inspecting pages. Its analysis presets do not offer an exploitation-off option. For a trial, use an authorized, isolated environment where possible and agree the testing scope with its owner. A subscription or login to someone else’s service is not permission to test it.

Will your code stay on your servers?

Separate two decisions: who hosts Vigilo, and who processes its model requests. The sample backend configuration lists Claude Code with Anthropic, AWS Bedrock or Azure Foundry; Codex with configurable endpoints; and OpenCode through LiteLLM for local or self-hosted models. These are documented connection options, not evidence that every model performs equally well.

A local gateway is not itself a local model. For example, LiteLLM’s Ollama integration requires a running model server. Trace the gateway’s actual destination, then check what code, prompts and tool results leave your environment. Approve the chosen provider’s access, retention and regional terms separately; a configuration example does not settle those questions.

The distinction also appears in Claude Managed Agents self-hosting, a separate product where hosting the agent service does not mean hosting its model. Its commercial terms do not apply to Vigilo.

There is operating work beyond choosing a model. Vigilo’s deployment reference describes persistent job storage, PostgreSQL and a workflow service called Temporal. Budget for access control, updates, storage and backups. The README also warns that the platform can start without a usable model backend, but scans cannot run.

What does a Vigilo trial cost?

The EUPL-1.2 licence grants royalty-free software rights, not free model processing or hosting. Build a budget with four parts: model usage, infrastructure, operation and expert review.

For a limited illustration, the configuration sample names Claude Opus 4.7 as one model option, not a universal default. As checked on 17 November 2026, Anthropic’s first-party pricing table lists it at US$5 per million base-input tokens and US$25 per million output tokens. Tokens are the units used to meter model input and output.

Illustrative calculation: 1 million uncached input tokens plus 100,000 output tokens would cost (1 × $5) + (0.1 × $25) = $7.50. This prices a hypothetical token bundle, not a typical Vigilo scan. It assumes standard first-party processing without caching, special modes, regional premiums, tool charges or discounts, and excludes tax.

Actual spending depends on the models and token categories used, repeated work and retries. Cache writes and reads have separate rates. Server costs and the time spent assessing findings remain outside that token subtotal; the available evidence does not establish an average scan cost or a saving.

Do not mistake every displayed dollar total for an invoice. The sample configuration describes Codex cost totals as estimates based on configurable prices. Reconcile recorded usage with the selected provider’s billing categories and current rates.

What does the licence allow?

The licence text permits use, modification and distribution subject to conditions. These include retaining notices, marking modifications, providing source when distributing or communicating the work, and copyleft rules—requirements to preserve the relevant licensing terms—subject to compatibility provisions.

Its definition of distribution or communication includes giving another person access to essential functionality. Do not assume a hosted service escapes those conditions. If you plan to redistribute a modified build or offer Vigilo to external users, have that arrangement reviewed against the licence. The software comes without general warranties, and its availability does not imply CERT-EU endorsement of a derivative service.

When should your team proceed—or wait?

CERT-EU says code owners must validate both findings and proposed patches before applying fixes. Vigilo’s automated critique is part of its workflow, not a substitute for that responsibility.

Use the following as a proposed trial decision, not a tested performance checklist:

  • Proceed with a bounded trial if you have a defined application and authorized scope, an approved model-data route, someone to operate the system and named reviewers. Set a spending limit and a decision date.

  • Wait if supported production operation, established detection accuracy or verified offline processing is a hard requirement you cannot yet satisfy. A browser-only team without source access also does not fit the documented current route.

  • Before expanding, compare useful, owner-validated findings and acceptable fixes with model spending and reviewer effort. Retain the exact software, model and configuration details alongside the results. A high finding count alone is not a success measure.

Name the person who can approve the next commitment; the related guide to deciding when to stop funding an AI pilot explains that distinction. The immediate decision is whether Vigilo earns a place in your existing review process—not whether it can replace that process on the strength of its release announcement.