The infrastructure
AI agents run on
Event-driven, durable, typed pipelines — exactly what production AI workflows need. LLM nodes, MCP tool discovery, multi-agent fan-out, and full observability. All in FlowDSL.
Why event-driven beats monolithic AI frameworks
Most AI agent frameworks are single-process orchestrators. Redelay is infrastructure — each agent step is an independent, resumable, auditable event.
MCP Tool Discovery
Every module operation is auto-exposed as an MCP tool at POST /mcp. AI assistants like Claude discover and invoke your entire API without any integration code. The mcp module is a blank-import.
LLM as Pipeline Node
FlowDSL's kind: llm treats language models as first-class pipeline stages — not afterthoughts bolted on. Declare scoring, classification, summarization, and routing decisions in YAML.
Durable by Default
Agent steps use durable delivery: MongoDB-persisted packets, exponential retry, idempotency keys, and checkpoint-based recovery. LLM timeouts and partial failures don't lose work.
Multi-Agent Fan-out
One agent's output event triggers multiple downstream consumers simultaneously. Coordinate specialized agents — researcher, validator, scorer, notifier — with zero tight coupling.
Typed Event Contracts
Every agent output is a schema-validated typed event. AsyncAPI generated automatically from your FlowDSL. Hallucinated or malformed LLM outputs are rejected before propagating downstream.
Full Observability
Every event carries a correlation ID, actor, and timestamp — threaded end-to-end through the entire agent chain. Replay any step, audit decisions, and debug agent behavior with ClickHouse.
Zero-config MCP tool server
Blank-import the mcp module. Every operation in every loaded module is instantly available to any MCP-compatible AI client — Claude, Cursor, Copilot, or your own agent.
// Blank-import to enable MCP endpoint
import (
_ "github.com/redelay/go-framework/modules/mcp"
_ "github.com/redelay/go-framework/modules/auth"
_ "github.com/redelay/go-framework/modules/users"
)
// POST /mcp now exposes:
// tools/list → auth.login, users.create,
// users.get, users.list, ...
// tools/call → invoke any operation
func main() {
inst, _ := app.Bootstrap()
server.Default(inst).Run()
}Real-world agent pipelines
Complete FlowDSL definitions for production AI workflows. Each example shows the delivery mode choice and why it matters.
Content Moderation Agent
llmrouterdurableAI-powered moderation pipeline: user-submitted content flows into an LLM analyzer, routed by confidence score to auto-approve, human review, or auto-reject. Durable delivery ensures no content is lost if the LLM times out.
Why Redelay makes this easier
- durable delivery means LLM timeouts never lose work
- checkpoint mode makes human review resumable after restarts
- typed ContentAnalysis prevents malformed LLM output from propagating
Start building AI-powered workflows
The getting started guide has you running a Go service with Kafka, auth, and events in under 5 minutes. Add FlowDSL and an LLM node when you're ready to go further.