⚠️ DEPRECATED — this project is no longer developed

SMAGo was a Go-based self-modifying AI agent. It has been superseded by SMAXr, the Elixir rewrite.

SMAXr keeps the self-modify idea, swaps Go for the BEAM, adds hot-reload of .ex files via apply_patch, an eval tool that compiles defmodule blocks live, an async turn loop with /stop and queued-message auto-inject, and a clean .env-based setup.

👉 Go here: github.com/AsmanovLev/SMAXr

A lightweight Go agent that communicates via Telegram, calls any OpenAI-compatible LLM, stores conversation history in SQLite, and can modify its own source code and binary at runtime.

SMAGo runs as a Windows system tray application with a supervisor that auto-restarts on crash and manages hot-swap upgrades.

Written entirely by AI. Minimax M3 and DeepSeek via OpenCode created the architecture and initial boilerplate. Then SMAGo via mimo-v2.5 (vision) developed itself. Human role: idea, direction, code review.

🇷🇺 Читать на русском: README.ru.md

• Telegram bot — long-polling, no webhooks needed
• Multi-provider LLM — OpenCode, DeepSeek, llama.cpp, or any OpenAI-compatible API
• Tool calling — terminal, read/write/edit files, web search, vision
• Self-modification — upgrade, rollback, restart via self_modify tool or Telegram commands
• Supervisor — system tray icon, auto-restart on crash, version swap with bad-version detection
• Markdown rendering — headings as bold, tables with alignment, code blocks
• Typing indicator — bot shows "typing..." while processing
• Session management — SQLite-backed, multi-session, per-chat
• DCP — Dynamic Context Pruning to stay within model context windows
• Stop/abort — interrupt long-running tasks gracefully or forcefully

• Windows 10/11
• Go 1.26+ (download)
• Git
• Telegram bot token — get one from @BotFather

winget install --id GoLang.Go -e --source winget
winget install --id Git.Git -e --source winget

Or download manually from go.dev and git-scm.com.

git clone git@github.com:AsmanovLev/SMAGo.git
cd SMAGo

build.bat

This produces three binaries in bin/:

copy config.example.json config.json

Edit config.json and fill in:

1. Start a chat with your bot in Telegram
2. Send any message
3. Open data/smago.log and look for the chat ID, or use @userinfobot

Keep sensitive values out of config.json:

set SMAGO_TELEGRAM_TOKEN=your_token
set SMAGO_TELEGRAM_CHAT_ID=your_chat_id
set SMAGO_OPENCODE_KEY=your_api_key

With supervisor (recommended):

start-supervised.bat

The supervisor runs silently in the system tray. Right-click the tray icon for options.

Without supervisor:

bin\smago-bg.exe

Debug mode (console output):

bin\agent.exe

Send /start to your bot in Telegram. You should see a help message.

| self_modify | Restart, upgrade, rollback, or check version | | my-server__* | Dynamic tools from MCP servers (see below) |

SMAGo supports external tool servers via MCP. Any stdio-based MCP server works — Node.js, Python, Go, Rust.

1. You add an MCP server to config.json
2. SMAGo spawns the server process on startup
3. Tools from the server appear as <server>__<tool> (e.g. fs__read_file)
4. The LLM can call them just like built-in tools

1. Put your MCP server files in mcps/ (gitignored):

mkdir mcps\my-server

2. Add to config.json:

{
  "mcp": {
    "fs": {
      "command": ["npx", "-y", "@modelcontextprotocol/server-filesystem", "C:\\path\\to\\folder"],
      "enabled": true
    }
  }
}

3. Restart SMAGo. New tools appear in /tools.

• Max 10 tools per server (most important ones, in declaration order)
• Server logs go to data/smago.log
• Disabled servers are skipped entirely
• See mcps/README.md for more examples

SMAGo/
├── src/
│   ├── main.go                 # Entry point, signal handling, PID, logging
│   ├── config.go               # JSON config with multi-provider support
│   ├── llm.go                  # OpenAI-compatible chat completions client
│   ├── telegram.go             # Long-polling Bot API (stdlib only, zero deps)
│   ├── session.go              # SQLite store (modernc.org/sqlite, no CGO)
│   ├── agent.go                # Main loop: msg → LLM → tools → response
│   ├── tools.go                # Tool registry
│   ├── self_modify_tool.go     # Self-modification: upgrade, rollback, restart
│   ├── markdown.go             # Markdown → Telegram HTML
│   ├── dcp.go                  # Dynamic Context Pruning
│   ├── dcp_compress.go         # Context compression logic
│   ├── dcp_strategies.go       # Pruning strategies
│   ├── vision.go               # Image analysis via multimodal model
│   ├── web_search_tool.go      # DuckDuckGo HTML search
│   ├── shell.go                # Shell command execution
│   ├── http.go                 # HTTP client
│   ├── inject.go               # Prompt injection helpers
│   ├── git.go                  # Git operations for self-upgrade
│   ├── cmd_upgrade.go          # Upgrade build logic
│   ├── cmd/
│   │   ├── supervisor/         # System tray supervisor
│   │   └── genicon/            # Icon generation tool
│   └── go.mod
├── bin/                        # Built binaries
├── data/                       # Runtime data (sessions, logs, versions)
├── config.json                 # Your configuration (not in git)
├── config.example.json         # Example configuration
└── build.bat                   # Build script

SMAGo uses git commit SHAs as version identifiers:

data/
  versions/
    cff3262/agent.exe
    addc8d7/agent.exe
  current.json    → {"version": "cff3262"}
  next.json       → {"version": "addc8d7"} (pending swap)

The supervisor watches for next.json and swaps binaries gracefully. If a version crashes within 20 seconds, it is marked as bad and will not be used again.

1. Path passed as first CLI argument
2. $SMAGO_CONFIG environment variable
3. <exe-dir>\config.json
4. <cwd>\config.json
5. ~/.config/smago/config.json

The project started on June 16, 2026 and evolved rapidly over a few days.

The initial working prototype, written in a single session by Minimax M3 and DeepSeek:

• Telegram bot via long-polling (stdlib net/http, zero external deps)
• LLM — OpenAI-compatible chat completions (talked to local llama.cpp, self-hosted endpoints, OpenCode)
• 4 tools — bash (shell), read_file, write_file, list_dir
• Vision — image analysis via multimodal model (mimo-v2.5)
• SQLite sessions — per-chat conversation history (modernc.org/sqlite, no CGO)
• Markdown → HTML — headings, bold, italic, code blocks for Telegram
• Self-modification — agent could edit its own Go source, recompile, and swap the binary
• Supervisor — system tray app with auto-restart on crash
• Single binary — no Docker, no WSL, no external services

Providers were migrated from the author's opencode.json: local llama.cpp, a home server, and a self-hosted DeepSeek proxy.

From this point on, SMAGo developed itself. Each feature, fix, and refactor was written by the agent editing its own Go source code and recompiling via self_modify. The OpenCode agent supervised the process. Three times SMAGo broke itself badly enough that OpenCode had to step in and restore the codebase.

What SMAGo built on its own:

• Git integration & self-upgrade (34ede0a) — git.go: read git history, show diffs, use commit SHAs as version identifiers
• Abort & tool-call traces (76cb51d — b8d628c) — /stop and /abort commands, compact single-line trace format
• Major refactor (83e9e0a) — switch from sequential version numbers to git commit SHAs, tree-style tool trace with annotations, silent notifications, supervisor /rebuild
• Multi-session management (9d94ad8) — multiple named sessions per chat, tool-call annotations, self-upgrade confirmation prompt
• DCP — Dynamic Context Pruning (3d081b5) — the biggest feature: /compress, pruning strategies (dedup, error purge, system nudge), auto-calculated limits based on model context window, retry on HTTP 503/502/429/500 with exponential backoff
• Session management polish (b8bc85a — c6eff72) — /rename with LLM auto-naming, /sessions, /switch, /delete, command whitelist during active tasks, rich /help

• Removed ~300 MB of binaries from git history (filter-repo)
• Removed opencode-ref submodule
• Added .gitignore for build artifacts, logs, databases
• Added README.md (EN) and README.ru.md (RU)
• Removed Playwright browser, MCP client, and Node.js dependency
• Final tool set: terminal, read_file, write_file, edit_file, list_dir, web_search, vision, compress, self_modify