distrocool
DistroCool Brief

What changed in AI. What you can build with it.

RSS ↗
← All editions

4 min read

Three free local AI tools that give your agents real superpowers

If you are building AI agents or automations, you do not need expensive proprietary cloud subscriptions to give your code real-world capabilities. This week's edition focuses on three practical, open-source tools you can install and run completely free on your own computer: automate browser workflows with Browser Use, expose custom agent capabilities with FastMCP, and extract guaranteed schema-validated JSON from local Ollama models.

Open-source automation

Control any web page with Browser Use and local LLMs

What changed

The open-source Browser Use project surpassed 120,000 GitHub stars, providing a Python library that allows AI agents to interact with live web pages through Playwright. Instead of relying on brittle CSS selectors or custom scraping scripts, the agent observes the visual and DOM accessibility tree, clicks interactive elements, types into form fields, and extracts structured data.

Why it matters

Automating browser workflows typically breaks whenever website layouts change. Browser Use lets developers automate repetitive data extraction, form submissions, and testing using vision-enabled or local models without maintaining fragile selector maps.

Try it

Install with pip install browser-use playwright and run playwright install. Create a quick script: import asyncio, initialize Agent(task='Go to news.ycombinator.com and extract top 3 titles', llm=ChatOllama(model='llama3.2')), and run agent.run(). Inspect the browser actions in real time.

Availability & cost: The core Python library is completely free and open source under the MIT license on GitHub. It runs locally with Ollama, llama.cpp, or any OpenAI-compatible API endpoint without requiring cloud credits.

Limits: Complex multi-step tasks require capable vision or reasoning models to avoid getting stuck in navigation loops. CAPTCHA challenges, bot walls, and heavy single-page apps may require explicit human intervention.

Original source

Agent tooling

Build Model Context Protocol tools in ten lines with FastMCP

What you can do

FastMCP 2.0 provides a high-level, Pythonic framework for creating Model Context Protocol (MCP) servers. By adding a single @mcp.tool() decorator to standard Python functions with type hints and docstrings, FastMCP automatically creates compliant JSON schemas, error handling, and standard transport endpoints for AI agents.

Why it matters

Instead of writing raw JSON-RPC schemas and handling socket transport manually, builders can expose local databases, shell scripts, and system APIs to tools like Claude Desktop, Cursor, Codex, and Hermes in minutes.

Try it

Install with pip install fastmcp. In server.py, write: from fastmcp import FastMCP; mcp = FastMCP('Demo'); @mcp.tool() def add(a: int, b: int) -> int: '''Add two numbers''' return a + b. Run fastmcp dev server.py to test the interactive inspector, then link it to your agent config.

Availability & cost: Free and open source under the Apache-2.0 license. Works on Linux, macOS, and Windows with Python 3.10+ without any paid subscriptions or proprietary runtimes.

Limits: FastMCP handles the protocol layer and validation, but tool execution safety is the developer's responsibility. Do not expose destructive filesystem, network, or database operations without strict validation and approval guards.

Original source

Local AI models

Force local Ollama models into strict Pydantic JSON schemas

What you can do

The official Ollama Python library supports structured outputs natively via format=MySchema.model_json_schema(). Instead of writing complex regex parsers or hoping a model outputs valid JSON without markdown fences, Ollama constrains the token sampling directly against your schema.

Why it matters

Building production agents requires predictable data contracts. Schema-constrained decoding eliminates JSON syntax errors, missing dictionary keys, and hallucinations in data extraction pipelines while running 100% offline on your own machine.

Try it

In Python: from ollama import chat; from pydantic import BaseModel; class Lead(BaseModel): name: str; email: str; score: int; res = chat(model='llama3.2', messages=[{'role': 'user', 'content': 'Extract lead: Jane Doe jane@example.com high priority'}], format=Lead.model_json_schema()); lead = Lead.model_validate_json(res.message.content); print(lead).

Availability & cost: Free and open source. Runs locally on CPU or GPU with Ollama installed. No API keys, credit cards, or external network requests needed.

Limits: Strict grammar-constrained sampling ensures valid JSON structure, but does not guarantee the factual correctness of the extracted content. Smaller 1B-3B models may produce empty or hallucinated values if the prompt lacks clear context.

Original source

4 more signals

Useful checks and updates. Guidance shows its check date; event dates identify actual announcements.

  • Selvedge preserves agent architecture decisions locally

    An open-source MIT Python CLI and local MCP server lets developers store and query architectural decisions and rejected paths across AI coding sessions without sending project context to third-party servers.

  • Images Upscale enlarges UI screenshots without watermarks

    A clean browser-based utility running Real-ESRGAN models allows creators to upscale small UI screenshots and product images by 2x or 4x with free daily credits and no mandatory account.

  • Watch out for fake AI agent installer repos on GitHub

    Security researchers warn of typosquatted repositories mimicking popular open-source AI tools that distribute infostealers via setup scripts. Always verify repository star counts, commit history, and maintainer identities before running pip or npm commands.

  • Verify token addresses before copying from crypto transaction histories

    MetaMask cautions users against copying recipient addresses from past wallet activity due to zero-value address poisoning attacks. Always verify full alphanumeric contract addresses from official project documentation.

Build something cool

Build a repeatable operating brief for your agent

Before deploying browser automation or tool-calling agents, document explicit boundaries, prohibited actions, error-recovery steps, and human-approval gates. The free DistroCool operating brief template helps keep automated runs reliable and safe.

Open the DistroCool workflow →

Was this edition useful?

Your choice helps us shape future editions. No account needed.

Keep exploring

Browse the Daily Drop →Try the linked workflow →Browse Brief editions →