A Go library for building AI agent applications. Provides server factories, LLM abstractions, workflow orchestration, and multi-runtime deployment support.
- 🏭 Server Factories - A2A and HTTP servers in 5 lines (saves ~475 lines per project)
- 🧠 Multi-Provider LLM - Gemini, Claude, OpenAI, xAI, Ollama via OmniLLM
- 🔀 Workflow Orchestration - Type-safe graph-based execution with Eino
- ☁️ Multi-Runtime Deployment - Kubernetes (Helm) or AWS AgentCore
- 🔒 VaultGuard Integration - Security-gated credential access
agentkit/
├── # Core (platform-agnostic)
├── a2a/ # A2A protocol server factory
├── agent/ # Base agent framework
├── config/ # Configuration management
├── deploy/ # Deployment provider system
├── http/ # HTTP client utilities
├── httpserver/ # HTTP server factory
├── llm/ # Multi-provider LLM abstraction
├── orchestration/ # Eino workflow orchestration
│
├── # Platform-specific
└── platforms/
├── agentcore/ # AWS Bedrock AgentCore runtime
└── kubernetes/ # Kubernetes + Helm deployment
go get github.com/plexusone/agentkitpackage main
import (
"context""github.com/plexusone/agentkit/a2a""github.com/plexusone/agentkit/agent""github.com/plexusone/agentkit/config""github.com/plexusone/agentkit/httpserver"
)
funcmain() {
ctx:=context.Background()
cfg:=config.LoadConfig()
// Create agentba, _:=agent.NewBaseAgent(cfg, "research-agent", 30)
researchAgent:=NewResearchAgent(ba, cfg)
// HTTP server - 5 lineshttpServer, _:=httpserver.NewBuilder("research-agent", 8001).
WithHandlerFunc("/research", researchAgent.HandleResearch).
WithDualModeLog().
Build()
// A2A server - 5 linesa2aServer, _:=a2a.NewServer(a2a.Config{
Agent: researchAgent.ADKAgent(),
Port: "9001",
Description: "Research agent for web search",
})
// Start serversa2aServer.StartAsync(ctx)
httpServer.Start()
}This replaces ~100 lines of boilerplate with ~15 lines.
import"github.com/plexusone/agentkit/a2a"// Create and start A2A serverserver, _:=a2a.NewServer(a2a.Config{
Agent: myAgent, // Google ADK agentPort: "9001", // Empty = random portDescription: "My agent",
})
server.Start(ctx) // Blocking// orserver.StartAsync(ctx) // Non-blocking// Useful methodsserver.URL() // "http://localhost:9001"server.AgentCardURL() // "http://localhost:9001/.well-known/agent.json"server.InvokeURL() // "http://localhost:9001/invoke"server.Stop(ctx) // Graceful shutdownimport"github.com/plexusone/agentkit/httpserver"// Config-basedserver, _:=httpserver.New(httpserver.Config{
Name: "my-agent",
Port: 8001,
HandlerFuncs: map[string]http.HandlerFunc{
"/process": agent.HandleProcess,
},
})
// Builder pattern (fluent API)server, _:=httpserver.NewBuilder("my-agent", 8001).
WithHandlerFunc("/research", agent.HandleResearch).
WithHandlerFunc("/synthesize", agent.HandleSynthesize).
WithHandler("/orchestrate", orchestration.NewHTTPHandler(exec)).
WithTimeouts(30*time.Second, 120*time.Second, 60*time.Second).
WithDualModeLog().
Build()
server.Start()import (
"github.com/plexusone/agentkit/agent""github.com/plexusone/agentkit/config"
)
cfg:=config.LoadConfig()
ba, err:=agent.NewBaseAgent(cfg, "my-agent", 30)
iferr!=nil {
log.Fatal(err)
}
deferba.Close()
// Utility methodscontent, err:=ba.FetchURL(ctx, url, maxSizeMB)
ba.LogInfo("message %s", arg)
ba.LogError("error %s", arg)ba, secCfg, err:=agent.NewBaseAgentSecure(ctx, "secure-agent", 30,
config.WithPolicy(nil), // Use default policy
)
iferr!=nil {
log.Fatalf("Security check failed: %v", err)
}
deferba.Close()
defersecCfg.Close()
log.Printf("Security score: %d", secCfg.SecurityResult().Score)import (
"github.com/cloudwego/eino/compose""github.com/plexusone/agentkit/orchestration"
)
// Build workflow graphbuilder:= orchestration.NewGraphBuilder[*Input, *Output]("my-workflow")
graph:=builder.Graph()
// Add nodes using Eino's InvokableLambdaprocessLambda:=compose.InvokableLambda(processFunc)
graph.AddLambdaNode("process", processLambda)
formatLambda:=compose.InvokableLambda(formatFunc)
graph.AddLambdaNode("format", formatLambda)
// Connect nodesbuilder.AddStartEdge("process")
builder.AddEdge("process", "format")
builder.AddEndEdge("format")
// ExecutefinalGraph:=builder.Build()
executor:=orchestration.NewExecutor(finalGraph, "my-workflow")
result, err:=executor.Execute(ctx, input)
// Expose as HTTP handlerhandler:=orchestration.NewHTTPHandler(executor)
http.Handle("/execute", handler)AgentKit supports two deployment runtimes:
| Aspect | Kubernetes | AWS AgentCore |
|---|---|---|
| Distributions | EKS, GKE, AKS, Minikube, kind | AWS only |
| Config tool | Helm | CDK / Terraform |
| Scaling | HPA | Automatic |
| Isolation | Containers | Firecracker microVMs |
| Pricing | Always-on | Pay-per-use |
import"github.com/plexusone/agentkit/platforms/kubernetes"// Load and validate Helm valuesvalues, errs:=kubernetes.LoadAndValidate("values.yaml")
// Merge base and overlay valuesvalues, err:=kubernetes.LoadAndMerge("values.yaml", "values-prod.yaml")Example values.yaml:
global:
image:
registry: ghcr.io/myorgpullPolicy: IfNotPresenttag: "latest"namespace:
create: truename: my-agentsllm:
provider: geminigeminiModel: "gemini-2.0-flash-exp"agents:
research:
enabled: truereplicaCount: 1image:
repository: my-research-agentservice:
type: ClusterIPport: 8001a2aPort: 9001resources:
requests:
cpu: 100mmemory: 128Mivaultguard:
enabled: trueminSecurityScore: 50import"github.com/plexusone/agentkit/platforms/agentcore"// Simple setupserver:=agentcore.NewBuilder().
WithPort(8080).
WithAgent(researchAgent).
WithAgent(synthesisAgent).
WithDefaultAgent("research").
MustBuild(ctx)
server.Start()Wrap Eino executors for AgentCore:
// Build Eino workflowgraph:=buildOrchestrationGraph()
executor:=orchestration.NewExecutor(graph, "stats-workflow")
// Wrap for AgentCoreagent:=agentcore.WrapExecutor("stats", executor)
// Or with custom I/O transformationagent:=agentcore.WrapExecutorWithPrompt("stats", executor,
func(promptstring) StatsReq { returnStatsReq{Topic: prompt} },
func(outStatsResp) string { returnout.Summary },
)// Agent implementation - runtime agnosticexecutor:=orchestration.NewExecutor(graph, "stats")
// Runtime 1: KuberneteshttpServer, _:=httpserver.NewBuilder("stats", 8001).
WithHandler("/stats", orchestration.NewHTTPHandler(executor)).
Build()
// Runtime 2: AWS AgentCoreacServer:=agentcore.NewBuilder().
WithAgent(agentcore.WrapExecutor("stats", executor)).
MustBuild(ctx)AgentCore code runs locally without AWS - same binary, different infrastructure:
go run main.go
curl localhost:8080/ping
curl -X POST localhost:8080/invocations -d '{"prompt":"test"}'| Aspect | Local | AWS AgentCore |
|---|---|---|
| Process | Go binary | Firecracker microVM |
| Sessions | In-memory | Isolated per microVM |
| Scaling | Manual | Automatic |
No code changes needed between local development and production.
A2A (Agent-to-Agent) protocol server factory.
server, _:=a2a.NewServer(a2a.Config{
Agent: myAgent,
Port: "9001",
Description: "My agent",
InvokePath: "/invoke", // Default: /invokeReadHeaderTimeout: 10*time.Second,
SessionService: customService, // Default: in-memory
})HTTP server factory with builder pattern.
server, _:=httpserver.NewBuilder("name", 8001).
WithHandlerFunc("/path", handlerFunc).
WithHandler("/path2", handler).
WithTimeouts(read, write, idle).
WithDualModeLog().
Build()Base agent implementation with LLM integration.
ba, err:=agent.NewBaseAgent(cfg, "name", timeoutSec)
ba, secCfg, err:=agent.NewBaseAgentSecure(ctx, "name", timeout, opts...)Configuration management with VaultGuard integration.
Deployment provider system for multi-cloud container deployments.
import (
"github.com/plexusone/agentkit/deploy"
_ "github.com/plexusone/agentkit-aws-pulumi/deploy/providers/lightsail"
)
cfg, _:=deploy.LoadDeployConfig("deploy.yaml")
provider, _:=deploy.GetProvider(cfg) // Respects AGENTKIT_DEPLOY_PROVIDERdeferprovider.Close()
status, _:=provider.Deploy(ctx, cfg)
fmt.Println(status.Outputs["serviceUrl"])cfg:=config.LoadConfig()
secCfg, err:=config.LoadSecureConfig(ctx, config.WithDevPolicy())
apiKey, err:=secCfg.GetCredential(ctx, "API_KEY")LLM model factory and adapters.
factory:=llm.NewModelFactory(cfg)
model, err:=factory.CreateModel(ctx)Eino-based workflow orchestration.
builder:= orchestration.NewGraphBuilder[Input, Output]("name")
executor:=orchestration.NewExecutor(graph, "name")
handler:=orchestration.NewHTTPHandler(executor)HTTP client utilities for inter-agent communication.
err:=http.PostJSON(ctx, client, url, request, &response)
err:=http.GetJSON(ctx, client, url, &response)
err:=http.HealthCheck(ctx, client, baseURL)Helm chart value structs and validation for Kubernetes deployments.
values, errs:=kubernetes.LoadAndValidate("values.yaml")
values, err:=kubernetes.LoadAndMerge("values.yaml", "values-prod.yaml")AWS Bedrock AgentCore runtime support.
server:=agentcore.NewBuilder().
WithAgent(agent).
MustBuild(ctx)
// Wrap Eino executorsagent:=agentcore.WrapExecutor("name", executor)AgentKit loads configuration from environment variables:
| Variable | Description | Default |
|---|---|---|
LLM_PROVIDER | LLM provider (gemini, claude, openai, xai, ollama) | gemini |
LLM_MODEL | Model name | Provider default |
GEMINI_API_KEY | Gemini API key | - |
CLAUDE_API_KEY | Claude/Anthropic API key | - |
OPENAI_API_KEY | OpenAI API key | - |
XAI_API_KEY | xAI API key | - |
OLLAMA_URL | Ollama server URL | http://localhost:11434 |
OBSERVABILITY_ENABLED | Enable LLM observability | false |
OBSERVABILITY_PROVIDER | Provider (opik, langfuse, phoenix) | opik |
AgentKit eliminates ~1,500 lines of boilerplate per project:
| Component | Lines Saved |
|---|---|
| A2A server factory | ~350 lines |
| HTTP server factory | ~125 lines |
| Shared pkg/ code | ~930 lines |
See BENEFITS.md for detailed analysis.
AgentKit has companion modules for Infrastructure-as-Code (IaC) deployment:
| Module | Purpose | Dependencies |
|---|---|---|
| agentkit-aws-cdk | AWS CDK constructs for AgentCore | 21 |
| agentkit-aws-pulumi | Pulumi components for AWS Lightsail | 340 |
| agentkit-k8s-pulumi | Pulumi components for Kubernetes (EKS/GKE/AKS) | - |
All modules share the same YAML/JSON configuration schema from platforms/agentcore/iac/.
For pure CloudFormation (no CDK/Pulumi runtime), use the built-in generator:
import"github.com/plexusone/agentkit/platforms/agentcore/iac"config, _:=iac.LoadStackConfigFromFile("config.yaml")
iac.GenerateCloudFormationFile(config, "template.yaml")See ROADMAP.md for planned modules including Terraform support.
- OmniLLM - Multi-provider LLM abstraction
- VaultGuard - Security-gated credentials
- Eino - Graph-based orchestration
- Google ADK - Agent Development Kit
- a2a-go - A2A protocol implementation
MIT License