Celladore Sluice
Celladore Sluice gives applications one governed boundary for model and API traffic. It authenticates callers, mediates supported protocols, selects an allowed provider, applies fail-closed policy, and records the routing decision without taking ownership of agent execution.
npx @celladore/sluice init- OpenAI-compatible boundary
- supported clients change their base URL and retain familiar request shapes
- Per-consumer virtual keys
- allowlists, budgets, attribution, and a central kill switch
- Inspectable decisions
- safe alpha views explain representative routing and policy outcomes
How Celladore Sluice Connects to the Stack
Celladore is engineered as a unified, composable toolchain where every tool feeds upstream contexts and downstream verification loops without friction.
→ Celladore Docket
Sluice emits OpenTelemetry spans and Azure OpenAI diagnostic events that Docket ingests for per-project spend attribution.
→ Retort
All Retort agent LLM calls route through Sluice, gaining failover, rate limits, and usage attribution without changing a single line of agent code.
→ Deck
Deck reads Sluice's unauthenticated /metrics endpoint directly for gateway health tiles and throughput sparklines in the Dashboard panel.
Technical Matrix
- Runtime & Engine:
- Python 3.12 / LiteLLM Proxy / FastAPI
- Deployment Target:
- Azure Container Apps (Dynamic Autoscale)
- Communication Protocol:
- OpenAI REST (/v1/responses, /v1/embeddings, /v1/audio/transcriptions)
- Licensing Tier:
- Proprietary
- Release Stage:
- Public Alpha
Invented from Real Scars
Rogue storytelling agent loops racked up thousands in token bills in a single week with zero budget visibility and fatal provider outages disrupting live inference.
Provider downtime, unmetered API key leakage, and runaway recursive model loops.
Architectural Features & Guarantees
Modular design built for high velocity, deterministic reliability, and strict enterprise isolation.
Centralized Auth & Credential Isolation
One shared gateway key replaces Azure OpenAI credentials scattered across every consuming project. A single revocation kills access everywhere instantly.
Rate & Budget Governance
Per-key RPM, TPM, and spend caps enforced at the gateway with a kill switch that doesn't require changes to any consumer code.
Built-in Observability
Prometheus metrics, OpenTelemetry tracing to Application Insights, structured Log Analytics logs, and optional Langfuse LLM observability — without instrumenting callers.
Semantic Response Caching
Optional Redis-backed cache cuts repeat-prompt cost and latency. Identical or semantically near-identical prompts return cached responses transparently.
Per-User Model Selection
FastAPI State Service lets each user select their preferred model from a governed catalog — backed by Azure Cache for Redis in production.
Multi-Environment Deployment
Independent dev, staging, and prod deployments on Azure Container Apps — each with its own Container Apps, Key Vault, and optional Redis instance.
Typed API Contract & Snippets
Live cURL, TypeScript, and Python invocation payloads for Celladore Sluice.
Unified Developer Interface
Route LLM completions through LiteLLM on Azure Container Apps with per-workspace virtual key metering and budget tagging.
import OpenAI from "openai";
const client = new OpenAI({
apiKey: process.env.CELLADORE_VIRTUAL_KEY || "YOUR_VIRTUAL_KEY",
baseURL: "https://litellm.sluice.celladoresystems.com/v1",
defaultHeaders: {
"X-Celladore-Budget-Tag": "project:celladore-web",
},
});
const response = await client.chat.completions.create({
model: "claude-3-7-sonnet-latest",
messages: [
{ role: "system", content: "You are an autonomous engineering agent." },
{ role: "user", content: "Verify PR quality gate requirements and summarize." },
],
temperature: 0.2,
});
console.log(response.choices[0].message.content);Status Notice: Celladore Sluice is an alpha. The public demo uses labelled sample decisions and never exposes credentials or production traffic. The Celladore endorsement is part of the public identity while standalone naming remains under review before beta or GA. Get in touch with the founder.