god-agent

Personal AI agent daemon accessible via Telegram. Built on Claude, with persistent memory, skill learning, and extensible tool orchestration. The agent's internal codename is Anima; that is the name it answers to.

Developed privately (181 commits, April-May 2026); history filtered from the private original; personal and unrelated material removed.

What It Does

• Answers questions, runs commands, manages files, searches the web
• Remembers across sessions (hybrid vector+keyword memory with dream consolidation)
• Learns new skills from successful task patterns (extraction, probation, routing feedback)
• Schedules recurring tasks (cron) and runs background compute during idle time
• Processes documents and images natively; audio/video workflows are CLI-backed and dependency-gated
• Browses websites and tests apps through local CLI/MCP tooling such as OpenCLI and Maestro when installed
• Runs multi-source social research through local tools such as last30days when installed
• Self-diagnoses via /doctor, gracefully interrupts on new messages

Architecture

Runtime-registered builtin tools, skills, 12-layer security, and 3-tier memory with dream synthesis. Optional CLI tools (ffmpeg, pandoc, gh, opencli, aria2c, tesseract, maestro, whisper, last30days) are invoked through bash when installed and are reported by /doctor as available or degraded. Agent loop with stuck detection, failure diagnosis, effort scaling, graceful user interrupt, and trajectory-based learning.

src/
  index.ts          # CLI entry + Telegram bot orchestration
  agent/            # Agent loop, effort scaling, stuck detection, verification, trajectory
  tools/            # Tool registry, builtins, MCP client, executor, document processing
  memory/           # Hybrid search, dream consolidation, decay, embeddings, integrity
  skills/           # Loader, extractor, versioning, probation, routing, quality gate, dedup
  security/         # 12 composable layers (injection, leak, redact, SSRF, Rule of Two)
  providers/        # Anthropic SDK + provider chain (retry, cache, cost guard, recording)
  telegram/         # Bot, streaming bridge, commands, access control, flood control
  autonomy/         # Cron scheduler, sleep-time compute, heartbeat
  sandbox/          # Docker container execution with network proxy
  config/           # Zod-validated YAML + env config
  db/               # SQLite wrapper (WAL, FTS5, sqlite-vec)
  resilience/       # Disk monitoring, auto-cleanup
  service/          # launchd/systemd process supervision
  workspace/        # Git-backed workspace initialization
  utils/            # Logger (pino), EventBus

Requirements

• Node.js >= 22
• pnpm >= 10
• Ollama (for embeddings)
• Docker (optional, for sandbox)

Setup

pnpm install
cp config.example.yaml config.yaml
# Edit config.yaml with your API keys, or set via env:
# GOD__LLM__API_KEYS, GOD__TELEGRAM__BOT_TOKEN, GOD__TELEGRAM__ALLOWED_USER_IDS

Usage

pnpm build
node dist/index.js start    # Start the daemon
pnpm dev                    # Development mode (tsx)

pnpm test                   # Run tests
pnpm typecheck              # Type check
pnpm lint                   # Lint

Configuration

YAML config with environment variable overrides (GOD__SECTION__KEY format):

GOD__LLM__API_KEYS="sk-ant-..."
GOD__TELEGRAM__BOT_TOKEN="..."
GOD__TELEGRAM__ALLOWED_USER_IDS="[123456]"
GOD__SEARCH__API_KEY="..."                      # Brave Search (optional, DDG fallback)
GOD__MEMORY__EMBEDDING__MODEL="qwen3-embedding:0.6b"

Telegram Commands

Command	What it does
/opus /sonnet /haiku	Switch model
/status	System status (model, cost, tools, tasks)
/doctor	Health diagnostics (14 checks)
/autonomy 0|1|2|3	Set autonomy level (supervised → full trust)
/effort low|medium|high|off	Override effort scaling
/dream	Trigger memory consolidation
/memory	Show memory contents
/recall on|off	Toggle active memory recall
/sandbox on|off	Toggle Docker sandbox
/approvals	List/remove tool approvals
/reset	Clear conversation history