# Pythinker: The Tiny AI Agent Framework Built for Real Builders Canonical: https://pythoughts.com/@melkholy/posts/post-1779509260256 Markdown: https://pythoughts.com/@melkholy/posts/post-1779509260256.md Author: Mohamed Elkholy (@melkholy) Published: 2026-05-23T04:11:08.260Z Updated: 2026-05-23T04:11:08.260Z > AI agents are becoming the next big shift in software. Not just chatbots. Not just wrappers around an API. Real agents that can remember, use tools, work across... ![image](/api/media/dbdf637a-d235-4684-b347-e5c2eb53a4b2/f5o27YCUdHfGxdNvq9NnG/uploads/1779509253553-nd23vi-image.png) AI agents are becoming the next big shift in software. Not just chatbots. Not just wrappers around an API. Real agents that can remember, use tools, work across apps, automate tasks, and help you build faster. But here is the problem: most AI agent frameworks are either too heavy, too complicated, or too locked into one ecosystem. **Pythinker** was built for people who want something different. It is an open-source, lightweight AI agent framework designed to help developers build personal agents that are powerful, flexible, and easy to understand. At its core, Pythinker keeps the agent loop compact and readable while still supporting chat channels, long-term memory, MCP tools, browser automation, subagents, and practical deployment paths. ## A Small Framework With Big Agent Energy Pythinker is intentionally tiny. That does not mean limited. It means focused. Instead of forcing you into a giant orchestration platform, Pythinker gives you a clean, readable foundation for building agents your way. You can inspect it, customize it, extend it, and actually understand what is happening under the hood. For developers, that matters. Because when your AI agent is doing real work, you do not want a black box. You want control. Pythinker gives you that control without making you fight the framework. ## One Agent, Many Channels Your AI agent should not live in only one place. Pythinker is channel-agnostic, meaning one Python process can listen across multiple communication channels, including Slack, Telegram, Discord, WhatsApp, Matrix, Microsoft Teams, email, a WebSocket WebUI, and an OpenAI-compatible HTTP API. That means you can build one agent and bring it into the places where you already work. Want a personal assistant in Telegram?A team helper in Slack?A support bot in Discord?A local developer agent in your terminal? Pythinker is built for all of that. ## Bring Your Favorite Models Pythinker is provider-rich by design. It supports 25+ LLM providers behind one interface, including Anthropic, OpenAI, Azure OpenAI, OpenAI Codex, GitHub Copilot, Qwen, MiniMax, VolcEngine, Moonshot, DeepSeek, StepFun, and more. This gives you freedom. You can choose the best model for coding, research, automation, cost, privacy, or speed. You are not locked into one provider. You can experiment, compare, and switch as your needs change. Even better, Pythinker supports provider hot-reload. When you edit your model, provider, or API key in the config file, the change lands at the next turn boundary without restarting everything. That makes iteration feel fast and natural. ## Built for Developers Who Actually Build Pythinker is not just a toy framework. It is designed for practical, daily use. You can start it from the CLI, launch a full-screen terminal chat, onboard a provider, manage sessions, switch models, inspect tasks, and run agents from a clean developer workflow. The full-screen TUI gives you live streaming chat, slash-command pickers, fuzzy search, themes, model switching, provider switching, session management, MCP status, login/logout flows, and more. In other words, it feels like an actual workspace for agents, not just a demo script. ## Agents That Remember A useful agent should not forget everything after every conversation. Pythinker includes a long-term memory system through a two-phase “Dream” process that consolidates memory into files like `MEMORY.md`, `SOUL.md`, and `USER.md`. This makes it possible to build agents that learn your preferences, remember useful context, and become more personalized over time. That is where AI agents start to feel less like generic tools and more like real companions for work. ## Tools, Skills, Browser Automation, and MCP Pythinker is built to connect. It supports bundled skills such as GitHub, cron, weather, tmux, summarization, and skill creation. It also includes first-class Model Context Protocol support, so you can connect external tools in a more structured way. Need browser automation? Pythinker includes an opt-in headless browser tool powered by Playwright-managed Chromium. It can handle JavaScript-rendered pages, click and form flows, screenshots, and DOM snapshots. That means your agent can do more than chat. It can interact with websites, use tools, gather information, automate workflows, and operate across real environments. ## Subagents for Bigger Tasks Sometimes one agent is not enough. Pythinker supports autonomous subagent tracking. You can spawn subagents as durable task records with saved output under `.pythinker/task-results/`. You can also choose roles such as `coder`, `explore`, or `plan`, depending on what kind of task you want the subagent to perform. From chat, you can use commands like: ```bash /tasks /task-output /task-stop ``` This makes Pythinker useful for longer-running work, research, coding tasks, planning, and background automation. ## Safer by Design AI agents need power, but they also need boundaries. Pythinker includes practical safety features such as sandboxed shell execution on Linux, workspace boundary enforcement for file tools, allow-list checks for channel ingress, and optional governed runtime controls. For hardened deployments, the governed runtime can include policy allow-lists, per-turn budgets, recursion depth controls, request context, telemetry sinks, and tool egress gateways. That makes Pythinker suitable not only for personal experiments, but also for more serious internal tools and production-minded workflows. ## What Can You Build With Pythinker? Pythinker gives you the foundation to build many types of agents, including: - Personal AI assistants that remember your preferences. - Developer agents that help with coding, debugging, and research. - Team bots for Slack, Discord, Telegram, or Microsoft Teams. - Research agents that browse, summarize, and generate reports. - Workflow automation agents for daily routines. - Knowledge assistants connected to your files, tools, and memory. - Long-running local agents that operate across multiple channels. The best part is that you are not forced into one pattern. Pythinker gives you the core building blocks, then lets you design the agent you actually want. ## Easy to Install and Start You can get started with the Python fallback install: ```bash pip install pythinker-ai ``` Then run: ```bash pythinker --version pythinker onboard pythinker ``` Pythinker also provides a quick-start workflow for onboarding, provider login, interactive CLI chat, and full-screen TUI usage. ## Why Pythinker Matters The future of AI agents will not be one giant app that everyone uses the same way. It will be personal.Customizable.Connected.Memory-aware.Tool-powered.And built around the way each person or team actually works. That is what Pythinker is aiming for. It gives developers a tiny but capable framework for building agents that are understandable, flexible, and practical. You can start simple, then grow into advanced workflows with memory, tools, browser automation, subagents, MCP, multiple providers, and real deployment options. ## Build Your Agent. Own Your Workflow. If you want an AI agent framework that is open-source, readable, provider-flexible, and built for real customization, Pythinker is worth exploring. Start with a local setup. Connect your favorite model. Add memory. Try the TUI. Bring it into your chat apps. Then shape it into the agent you always wanted. **Ready to build your own personal AI agent? Install Pythinker today and start creating agents that work where you work.**