Open Source

Our contribution
to the open AI ecosystem

In building aiClaw we create tools that are useful on their own. We share them with everyone.

2
Open Source projects
15
Built-in tools
8+
LLM providers supported
MIT
License

agent-runner

deploy-sh/agent-runner TypeScript · Node.js MIT

Universal agentic CLI — connect any LLM with OpenAI-compatible API and get a full agent with tools. Works as interactive REPL or in one-shot mode for scripts.

agent-runner gives any language model an agentic loop with tools: query → LLM thinks → calls tool → sees result → thinks again → answer. Just provide provider API key and model — everything else works out of the box.

  • Two modes: interactive REPL (like Claude Code) and one-shot run for scripts and automation
  • Support for models without native tool_calling via --fallback — prompt compatibility mode
  • MCP client: connect additional tools from any MCP server via SSE
  • Auto context compression on overflow — long sessions without interruptions
  • Vector memory: save and search facts between sessions (memory_write, memory_search)
  • FORKING.md — ready instructions for creating a branded fork for your own provider

Built-in tools (15)

bash read_file write_file edit_file list_dir grep http_request pdf_to_text youtube_transcript web_search python_exec spawn_agent memory_write memory_search agent_browser

Supported providers

OpenAI
Mistral AI
Groq
Any OpenAI-compat API
Install with one command
# Установить в /usr/local/bin
curl -fsSL https://raw.githubusercontent.com/
deploy-sh/agent-runner/main/install.sh | bash

# Первый запуск — настройка
agent-runner --setup

# Запустить REPL
agent-runner
Usage examples
# Интерактивная сессия
$ agent-runner
> проанализируй логи в /var/log/nginx

# Однократный запрос
$ agent-runner -p "найди утечки памяти"

# Конкретная модель
$ agent-runner --model qwen3:8b

giga-cli

deploy-sh/agent-runner · ветка giga_cli Форк agent-runner

Branded fork of agent-runner for GigaChat by Sber. Full agentic loop on a domestic LLM — with OAuth, compatibility adapters and all 15 tools.

GigaChat is an LLM from Sberbank with a free tier and Russian infrastructure. giga-cli adapts the agentic loop to the platform's specifics: OAuth authorization, tool name and argument normalization.

  • GigaChat OAuth out of the box — authorization via ClientID without manual token management
  • normalizeToolName() — GigaChat strips underscores: memorysearchmemory_search
  • normalizeArgs() — argument aliases for compatibility with different response formats
  • All 15 agent-runner tools — full functionality without limitations
  • GigaChat free tier works without balance — a good starting point

How the fork is structured

giga-cli differs from agent-runner only in identity and adapters

brand.ts Name, config directory (~/.giga-cli/), provider list
loop.ts GigaChat OAuth, normalizeToolName(), normalizeArgs()
tools.ts Same 15 tools, agent_browser via temp file
Config ~/.giga-cli/.env — GIGACHAT_CLIENT_ID + AGENT_MODEL
Config ~/.giga-cli/.env
# Бесплатная версия
AGENT_PROVIDER=gigachat
GIGACHAT_CLIENT_ID=ваш_client_id
AGENT_MODEL=GigaChat

# Pro версия
AGENT_MODEL=GigaChat-Pro
Build from source
# Клонировать ветку
git clone -b giga_cli \
https://github.com/deploy-sh/agent-runner

# Сборка бинаря
npx esbuild src/main.ts --bundle \
--platform=node --outfile=release/giga-cli.js
npx pkg release/giga-cli.js \
--target node22-linux-x64 \
--output release/giga-cli-linux-x64

Using giga-cli as a template, you can create a fork for any provider in an hour. Full instructions in FORKING.md.

Related solutions

Tools from the automation ecosystem that work well together with aiClaw.

repio.dev

External product

Record and replay macros — automation of another level

If an AI agent automates through understanding and reasoning, repio automates through exact action replay. Record a sequence of actions in the browser or app once, then run it as many times as you want.

The two approaches complement each other: aiClaw thinks and makes decisions, repio reproduces repetitive operations with click-level precision.

Record actions directly in the browser without code
Replay macros on schedule or by trigger
Work with forms, tables, interfaces without API
Fits where an AI agent is overkill
Go to repio.dev