Policy Enforcement
Budget gates, rate limits, and model access controls enforced at the proxy layer in real time. Every policy violation logged to an immutable audit trail.
Candela — LLM Governance Platform
See everything. Enforce anything. Open-source governance for your AI infrastructure — trace every request, enforce budgets, control model access, and prove compliance.
What You Get
Section titled “What You Get”Cost Governance
Real-time cost calculation per request, per model, per tenant. Set per-user budgets, create time-bound grants, and get threshold alerts before spend is exceeded.
Full Observability
Every LLM request captured with OpenTelemetry — latency, TTFB, token counts, cache hits, and cost — all in a unified trace tree with W3C Trace Context propagation.
Drop-In Integration
Works with Google ADK, LangChain, Claude Code, Cursor, Zed, OpenCode — anything that talks to an LLM endpoint. Point base_url at Candela and you’re done.
Governance That Ships Today
Section titled “Governance That Ships Today”Candela is not a roadmap — these enforcement controls are live in production:
Budget Enforcement
Per-user daily budgets with pre-flight gates. Requests that exceed limits get a 402 before they ever reach the LLM provider.
Rate Limiting
Per-user request throttling prevents runaway automation from draining budgets. Minute-window counters with configurable limits.
Immutable Audit Trail
Every request, every admin action, every policy decision — logged to BigQuery, DuckDB, or SQLite. Tamper-proof by design.
Tenant Isolation
Full multi-tenant attribution via X-Candela-Tenant-Id headers and W3C Baggage. Per-tenant cost tracking, leaderboards, and scoped views.
eBPF Enforcement
Kernel-level traffic interception via iptables redirect, Tetragon process enforcement, and Hubble observability. Zero SDK changes — every LLM request flows through the proxy, guaranteed.
Choose Your Component
Section titled “Choose Your Component” candela Developer proxy + runtime manager. Run local & cloud models with built-in governance and observability.
candela-server Full backend with dashboard UI, BigQuery storage, budget enforcement, policy controls, and team RBAC.
candela-sidecar Lightweight < 5MB proxy for production containers. Pub/Sub + OTLP export with budget enforcement.
candela-desktop Desktop app for provider management, cost monitoring, and trace visualization.
Enrichment SDKs Lightweight SDKs for Python, TypeScript, Go, Kotlin, and Rust — propagate tenant and job metadata for governance attribution.