🤖 gollem

• [x] Gemini (see models)
• [x] Anthropic Claude (see models)

• [x] OpenAI (see models)
• [x] Ollama (local or remote Ollama server, and Ollama Cloud; see LLM Provider Configuration)

Install

go get github.com/gollem-dev/gollem

Quick Start

package main
import (
	"context"
	"fmt"
	"os"
"github.com/gollem-dev/gollem"
	"github.com/gollem-dev/gollem/llm/openai"
)
func main() {
	ctx := context.Background()
// Create LLM client
	client, err := openai.New(ctx, os.Getenv("OPENAI_API_KEY"))
	if err != nil {
		panic(err)
	}
// Create session for one-time query
	session, err := client.NewSession(ctx)
	if err != nil {
		panic(err)
	}
// Generate content
	result, err := session.Generate(ctx, []gollem.Input{gollem.Text("Hello, how are you?")})
	if err != nil {
		panic(err)
	}
fmt.Println(result.Texts)
}

Features

Agent Framework

Build conversational agents with automatic session management and tool integration. Learn more →

agent := gollem.New(client,
	gollem.WithTools(&GreetingTool{}),
	gollem.WithSystemPrompt("You are a helpful assistant."),
)
// Session is managed automatically across calls
agent.Execute(ctx, "Hello!")
agent.Execute(ctx, "What did I just say?") // remembers context

Tool Integration

Define custom tools for LLMs to call, or connect external tools via MCP. Tools → | MCP →

// Custom tool - implement Spec() and Run()
type SearchTool struct{}
func (t *SearchTool) Spec() gollem.ToolSpec {
	return gollem.ToolSpec{
		Name:        "search",
		Description: "Search the database",
		Parameters:  map[string]*gollem.Parameter{
			"query": {Type: gollem.TypeString, Description: "Search query"},
		},
	}
}
func (t *SearchTool) Run(ctx context.Context, args map[string]any) (map[string]any, error) {
	return map[string]any{"results": doSearch(args["query"].(string))}, nil
}
// MCP server - connect external tool servers
mcpClient, _ := mcp.NewStdio(ctx, "./mcp-server", []string{})
agent := gollem.New(client,
	gollem.WithTools(&SearchTool{}),
	gollem.WithToolSets(mcpClient),
)

Multimodal Input

Send images and PDFs alongside text prompts. Learn more →

img, _ := gollem.NewImage(imageBytes)
pdf, _ := gollem.NewPDFFromReader(file)
result, _ := session.Generate(ctx, []gollem.Input{img, pdf, gollem.Text("Describe these.")})

Structured Output

Constrain LLM responses to a JSON Schema. Learn more →

schema, _ := gollem.ToSchema(UserProfile{})
session, _ := client.NewSession(ctx,
	gollem.WithSessionContentType(gollem.ContentTypeJSON),
	gollem.WithSessionResponseSchema(schema),
)
resp, _ := session.Generate(ctx, []gollem.Input{gollem.Text("Extract: John, 30, [email protected]")})
// resp.Texts[0] is valid JSON matching the schema

For one-shot queries, Query[T]() combines schema generation, session creation, LLM call, and JSON parsing into a single generic function call with automatic retry on parse failures:

type UserProfile struct {
	Name  string json:"name" description:"User's full name"
	Age   int    json:"age" description:"Age in years"
	Email string json:"email" description:"Email address"
}
result, _ := gollem.QueryUserProfile,
)
// result.Data is *UserProfile — type-safe, already parsed

To run a structured query on an existing session (preserving conversation history), use SessionQuery[T]():

// session already has conversation context from prior Generate calls
resp, _ := gollem.SessionQueryUserProfile
// resp.Data is *UserProfile, parsed from the LLM's JSON response
// The session's history (including this exchange) is preserved

Each provider receives the schema through its own schema parameter. For Claude models that support structured outputs, the schema is sent as output_config.format, and the system prompt and tool list are the same as in a call without a schema; older Claude models receive the schema in the system prompt (details). After a tool loop, WithToolCallsDisabled() keeps the tools in the request but forbids calling them, so a structured answer can be requested over the same history:

resp, _ := session.Generate(ctx, []gollem.Input{gollem.Text("Report the result as JSON.")},
	gollem.WithToolCallsDisabled(),
	gollem.WithGenerateResponseSchema(schema),
)

Middleware

Monitor, log, and control agent behavior with composable middleware. Learn more →

agent := gollem.New(client,
	gollem.WithToolMiddleware(func(next gollem.ToolHandler) gollem.ToolHandler {
		return func(ctx context.Context, req gollem.ToolExecRequest) (gollem.ToolExecResponse, error) {
			log.Printf("Tool called: %s", req.Tool.Name)
			return next(ctx, req)
		}
	}),
)

Strategy Pattern

Swap execution strategies: simple, ReAct, or Plan & Execute. Learn more →

import "github.com/gollem-dev/gollem/strategy/planexec"
agent := gollem.New(client,
	gollem.WithStrategy(planexec.New(client)),
	gollem.WithTools(&SearchTool{}, &AnalysisTool{}),
)

Tracing

Observe agent execution with pluggable backends (in-memory, OpenTelemetry). Learn more →

import "github.com/gollem-dev/gollem/trace"
rec := trace.New(trace.WithRepository(trace.NewFileRepository("./traces")))
agent := gollem.New(client, gollem.WithTrace(rec))

History Management

Portable conversation history for stateless/distributed applications. Learn more →

// Export history for persistence
history := agent.Session().History()
data, _ := json.Marshal(history)
// Restore in another process
var restored gollem.History
json.Unmarshal(data, &restored)
agent := gollem.New(client, gollem.WithHistory(&restored))

For automatic persistence, implement HistoryRepository and pass it via WithHistoryRepository. gollem then loads history at the start of a session and saves it after every LLM round-trip — no manual marshaling required.

agent := gollem.New(client,
    gollem.WithHistoryRepository(repo, "session-id"),
)
// History is loaded automatically on first Execute, and saved after each round-trip
err := agent.Execute(ctx, gollem.Text("Hello!"))

Examples

See the examples directory for complete working examples:

• Simple: Minimal example for getting started
• Query: Type-safe structured query with Query[T]()
• Basic: Simple agent with custom tools
• Chat: Interactive chat application
• MCP: Integration with MCP servers
• Tools: Custom tool development
• JSON Schema: Structured output with JSON Schema validation
• Embedding: Text embedding generation
• Tracing: Agent execution tracing with file persistence

Documentation

• Getting Started Guide
• Tool Development
• MCP Integration
• Structured Output with JSON Schema
• Middleware System
• Strategy Pattern
• Tracing
• History Management
• LLM Provider Configuration
• Debugging
• Compatibility Policy
• API Reference

License

Apache 2.0 License. See LICENSE for details.