Blog
Notes on smaller agent workflows.
Engineering, releases, and design tradeoffs for token-light local AI agents.
Recent posts
TinyWebUI: A Self-Hosted ChatGPT Alternative with Web Search and Memory Built In
Run npx tinywebui for a local AI chat UI that works with Ollama, LM Studio, vLLM, OpenRouter or OpenAI, with MCP web search and long-term memory already connected. No account, no Docker required.
TinySearch 0.6.4: Open-Source Browser Automation vs. ChatGPT Agent, Gemini, and Claude in Chrome
Every major closed-source lab now ships an agent that drives a browser for you. TinySearch 0.6.4 adds the same retrieval primitive, self-hosted, free, and deliberately more constrained.
What We Learned From Benchmarking Web Research Agents
We built an agent-in-the-loop benchmark to compare web research systems. The scores were useful, but the traces taught us much more about retrieval, context, latency, and tool design.
Link Discovery Without a Crawler: TinySearch 0.6.0
Scraping needs URLs you already have. Crawling follows links on its own budget. TinySearch 0.6.0 adds bounded related links so the agent picks the next page, plus support for an external browser over CDP.
TinySearch vs. Tavily, Exa, Perplexity Sonar, and Brave Search API
How TinySearch's self-hosted, no-key research pipeline compares to Tavily, Exa, Perplexity's Sonar API, and Brave's Search API for AI agents.
TinyContext vs. Mem0 and Zep: Memory for AI Agents Without a Second Database
Mem0 and Zep give agents durable memory, but both expect you to run a vector or graph database alongside them. TinyContext keeps memory in one SQLite file.
What Crawl4AI Does, and What TinySearch Adds on Top of It
Crawl4AI turns a known URL into clean markdown. TinySearch uses it for exactly that, then adds the search step, chunk ranking, and an MCP interface around it.
Add Web Search to a Local LLM: Open WebUI and TinySearch
Connect TinySearch to Open WebUI over MCP, choose a local model, and verify web search with source links. Includes setup and connection troubleshooting.
Why Raw Webpages Are Bad Context for AI Agents
Raw HTML is long, noisy, and expensive to put in a prompt. Agents need ranked evidence, not pasted webpages.
SearXNG vs DuckDuckGo: Differences, Setup and Agent Search
Compare SearXNG, DuckDuckGo and DDGS for browsing and AI agents. Learn what stays local, enable JSON search, and troubleshoot backend failures.
How to Reduce Token Usage in AI Agent Workflows
Reduce token usage in agent workflows by retrieving selectively, chunking aggressively, and treating context as a budget instead of a dump.
TinySearch: Open-Source MCP Web Research for Local LLMs
TinySearch is the first TinySuite tool: a self-hosted research layer for MCP agents that turns a question into cited, chunked evidence.