Guides
Task-focused how-to guides for every Crux capability, from the everyday building blocks up to durable runtime and operations.
Each guide starts from a problem you hit while building an AI feature and shows the Crux primitive that solves it. You do not need to read them in order. Jump to the one that matches what hurts right now.
New to Crux? Start with Foundations for the mental model, then Getting Started to ship your first prompt.
Building blocks
The core authoring primitives. Most projects start here.
Prompts
Typed, reusable prompt definitions with input/output schemas that run on any SDK.
Contexts
Reusable instruction fragments with conditional inclusion, priority, and caching.
Tools
LLM-callable functions declared on prompts and contexts, merged for the adapter.
Context & memory
Control what the model sees and what it remembers across turns.
Memory
Working state, episodes, facts, and procedures that persist across turns.
Retrieval & RAG
Load, index, sync, and retrieve documents so answers can cite sources.
Compaction
Keep long conversations under the context window without losing what matters.
Agents & workflows
Coordinate multiple model calls into agents, compositions, and resumable flows.
Agents
A prompt with an identity, default model, allowed tools, and handoffs.
Skills
Markdown instruction sets that agents load on demand.
Compositions
Pipeline, parallel, consensus, and swarm patterns over multiple agents.
Flows
Suspendable, resumable orchestration where each step is cached.
Background work
Run work after the response is sent, without blocking the user.
Plans & Tasks
Freeform plans and live task lists with reactive UI hooks.
Reliability
Make model choice, safety, and regressions explicit instead of hidden.
Routing & Fallback
Declare model routing, cascades, and fallback instead of hiding them in call sites.
Safety
Guardrails that block or redact I/O, and constraints that validate and retry.
Eval
Typed Cases, Variants, Baselines, and LLM-as-judge scoring for production AI behavior.
State & storage
Give runs a filesystem and give Crux somewhere to persist.
Workspaces
A versioned, transactional file surface agents and tools can read and write.
Storage
The store adapters behind memory, retrieval, plans, and runtime.
Runtime & operations
Make execution durable, then see and lint what happened.
Durable Execution
The optional engine that makes flows and background work survive restarts and retries.
Observability
Local devtools and OpenTelemetry to see exactly what the model saw and did.
Project Health
Static analysis and lints over your Crux definitions.
Project setup
Check and safely apply setup contributed by Crux subsystems with crux setup.