Risk 01
Per-developer model sprawl
Continue reads a local config.json. One developer uses Claude, another uses a self-hosted model, a third uses a free API key from an unknown provider. Security has no central view and no central control.
Continue is open source, pluggable, and loved by developers. It is also configured per-developer, which is the governance problem. Raidu gives you one runtime across every Continue install in your org.
Continue is configured through a local config.json. Every developer can pick models, endpoints, and prompts. Power for the developer. Visibility gap for security.
Continue's flexibility is its strength. Without a shared runtime, it is also four failure modes your security team will inherit.
Continue reads a local config.json. One developer uses Claude, another uses a self-hosted model, a third uses a free API key from an unknown provider. Security has no central view and no central control.
Continue sends file context and chat history to whichever endpoint the developer configured. Without central redaction, secrets and proprietary logic flow to providers your procurement team never approved.
If an incident happens, you cannot reconstruct which developer generated which code with which model under which policy. Continue writes no enterprise audit log.
Some developers use models with built-in safety. Others do not. Without a shared checkpoint layer, the org's risk is the weakest per-developer config, not the best.
Push a single config to every Continue install. Every developer hits Raidu. One policy, one audit trail, every endpoint choice.
Continue's config.json points at Raidu's OpenAI-compatible endpoint. You choose which models are available, enforce redaction before prompts leave, and flag any attempt to bypass the proxy.
Allowlist models per team, per repo, per file pattern. Claude for regulated code, GPT for internal tooling, self-hosted for air-gap. Developers pick from a curated menu. Security sets the menu.
Every completion is scanned for insecure code, license risks, hallucinated imports, and exfiltration before the developer sees it. Safe output reaches the editor. Blocked output leaves a logged reason.
Every Continue interaction is tied to developer identity, repo, policy version, and signed. RSA-4096 signed, SHA-256 chained, WORM retained. SOC 2 auditors can pull per-developer evidence on demand.
Continue reads its config from a local file. Ship the Raidu config via MDM, dotfiles repo, or onboarding script, and every install is governed.
{
"models": [
{
"title": "Claude Sonnet 4.5 (governed)",
"provider": "openai",
"model": "claude-sonnet-4.5",
"apiBase": "https://proxy.raidu.com/acme-corp/anthropic",
"apiKey": "raidu_xxx",
"requestOptions": {
"headers": { "x-raidu-policy": "coding.eng.v7" }
}
}
]
}
// Ship via dotfiles, MDM, or onboarding script. One source of truth.