$ today_in_10_seconds
Two open-source launches today address two different halves of the agent stack. Webcrawl gives any MCP-capable coding agent a local, keyless toolkit for reading the web. AgentRun puts typed workflows around your agents so the expensive model only runs when a cheap check fails.
$ lead_story
Webcrawl Gives Any MCP Agent Free, Local Web Access
Brainfish has open-sourced Webcrawl Anything, an MIT-licensed crawler. A single `npx webcrawl setup` installs it as an MCP server plus an Agent Skill across Claude Code, Cursor, Codex, Gemini and Windsurf. It runs locally with no API key and turns websites, JS-heavy SPAs, PDFs (with OCR), YouTube transcripts, GitHub repos and arXiv papers into clean Markdown. It also has structured extractors for CVEs (NVD, OSV and CISA KEV joined in one lookup), prices, travel and property listings. The detail worth copying is a compact schema profile, on by default, that shrinks its MCP tool definitions so they don't eat your agent's context window. It's open-core: the same engine powers a paid webcrawl.app tier with stealth browsers and caching, which puts it squarely against Firecrawl. Get legal sign-off before you point its anti-bot bypass tooling at anything in production.
// why_it_matters: Web access for agents is becoming a free commodity, so your moat has to be what your agent does with the page.
$ also_worth_attention
AgentRun: Typed Workflows That Call Your Agent Only When Needed
Parcha's AgentRun is a workflow DSL, written as JSON or with a TypeScript builder and Zod contracts, that wraps the agents you already run. Cheap deterministic steps run first. A 'judge' node then makes a typed yes/no/uncertain call with a confidence score. Only failures go to a full agent, and only after that to a human. In the bundled support demo, password resets and invoice lookups never invoke an agent at all. It's beta, and it's upfront about its limits: code nodes run with full process privileges, and judge calls need a TypeSafe API key.
link →
$ bigger_picture
The trend behind the headlines
Webcrawl expands what an agent can reach, and AgentRun constrains when and how it gets to act. Both ship as an MCP server or packaged skill installed directly into the coding agent a developer already uses, which shows where agent tooling now gets distributed. Both also follow the same open-core playbook: free open-source code with a paid service attached. Teams that pair cheap, broad capabilities with tight, auditable control flow will spend model calls only where judgment actually earns its cost.
Agent Skills specification
Both of today's launches package themselves as skills. Webcrawl installs a SKILL.md alongside its MCP server, and AgentRun's Pi extension ships one that teaches the agent to build workflows. If you're shipping a tool you want agents to use correctly, the spec for that format (a folder of instructions an agent loads on demand) is worth ten minutes.
$ subscribe --free
Get the next edition in your inbox
// DAIly Agent covers agentic AI three times a week
// free, no spam, exit anytime