saige

Super Artificial Intelligence Graph Environment
A Go SDK for building AI agents, giving them context and memory, and evaluating them.

Install · Report Bug · Go Docs

Features

saige focuses on three things: running agents, supplying their context and memory, and measuring both with evals.

Agents

• Streaming-first agent loop with typed delta events, parallel tool execution, sub-agent delegation, and handoffs
• 4 LLM providers (Ollama, OpenAI, Anthropic, Google) behind one Provider interface, with retry and fallback composition
• Nullable tool properties with separate presence rules across providers and MCP. See tool schemas.
• Durable runs that resume after a crash, on a local engine or on Postgres through duraturo, plus response caching
• MCP server exposing any saige tool pack to Claude Code, Codex, Gemini CLI, or any MCP client, with approval enforced for mutating tools
• Opt-in tool packs: workspace files (tools/fs), a sandboxed shell (tools/exec), and URL fetch with private-address blocking (tools/fetch). Read-only by default; every mutating tool requires approval
• HTTP and SSE server via saige serve: sessions, a resumable turn event stream in the versioned wire format, and approve and cancel endpoints
• Model catalog and presets as data: declared capabilities, layered JSON catalogs loaded from files, HTTPS or any reader, and presets whose failover entries each carry options validated for their own model. See model catalog and presets.
• MCP client with pooled sessions, safe retries, catalog drift detection, and .mcp.json loading. See MCP client.

Context and memory

• Conversation tree with branching, checkpoints, rewind, compaction, and RLHF feedback
• Multi-retriever RAG fusing vector, BM25, and graph retrieval via Reciprocal Rank Fusion, with reranking and citations
• Knowledge graph backend for RAG: LLM-powered entity extraction, fuzzy dedup, and temporal tracking. It differs from the document stores in ingestion and retrieval logic, not in role

Evals

• Composable scorers for agents, retrieval, and knowledge graphs, with gates, comparisons, experiments and LLM-as-judge
• Sampler to measure how stable scores and subjects are across repeated runs
• Live eval harness via saige eval

Why one SDK?

An agent is only as good as the context it is given, and you only know either works if you measure it. saige keeps all three under shared Provider, Embedder, and Tool interfaces, so retrieval plugs into the agent loop as tools and every layer is scored by the same eval framework.

Installation

Library

go get github.com/urmzd/saige

CLI and MCP server

go install github.com/urmzd/saige/cmd/saige@latest
go install github.com/urmzd/saige/cmd/saige-mcp@latest

Or install a pre-built binary (Linux and macOS, amd64 and arm64, checksum-verified):

curl -fsSL https://raw.githubusercontent.com/urmzd/saige/main/install.sh | bash
curl -fsSL https://raw.githubusercontent.com/urmzd/saige/main/install.sh | BIN=saige-mcp bash

Each release also attaches Windows amd64 builds and a SHA256SUMS file. Update an installed CLI with saige update (saige update --check only reports).

Quick Start

CLI

saige chat                                    # interactive multi-turn chat
saige ask "What is retrieval-augmented generation?"

# Serve the agent over HTTP + SSE with workspace file tools
saige serve --tools fs,fetch --workspace .

# Serve saige tools to Claude Code, Codex, or Gemini CLI over MCP
saige-mcp --tools all --db "$SAIGE_DB" --searxng-url http://localhost:8080

The CLI auto-detects a provider from ANTHROPIC_API_KEY, OPENAI_API_KEY, or GOOGLE_API_KEY, falling back to Ollama (no key needed). See the CLI reference for RAG/KG subcommands and flags.

Library

import (
    "github.com/urmzd/saige/agent"
    "github.com/urmzd/saige/agent/types"
    "github.com/urmzd/saige/agent/provider/ollama"
)

client := ollama.NewClient("http://localhost:11434", "qwen2.5", "nomic-embed-text")
a := agent.NewAgent(agent.AgentConfig{
    Name:         "assistant",
    SystemPrompt: "You are a helpful assistant.",
    Provider:     ollama.NewAdapter(client),
    Tools:        types.NewToolRegistry(myTool),
})

stream := a.Invoke(ctx, []types.Message{types.NewUserMessage("Hello!")})
for delta := range stream.Deltas() {
    switch d := delta.(type) {
    case types.TextContentDelta:
        fmt.Print(d.Content)
    }
}

See examples/ for runnable programs covering knowledge graphs, RAG pipelines, sub-agents, durability, and more.

Documentation

Each subsystem has its own README as the entrypoint for further information:

API reference for every package: pkg.go.dev/github.com/urmzd/saige

Agent Skill

This repo's conventions are available as portable agent skills in skills/.

Orchestration contracts

See ownership and policies for subagent results, handoff return links, sticky routing, and approval limits.
See cache contracts for cache identity, provider cache modes, and remaining defects.
See durable execution for saved approvals, crash recovery, and budget reservations.
See observability for OpenTelemetry spans, metrics, error attributes, and redaction.
See upgrade notes for behavior changes that can affect existing code.
Design decisions explain the choices and their limits.

License

Apache 2.0. See LICENSE.