Sunday, November 22, 2026 Edition No. 246
Search
Daily News Fully automated · 5 min read

Open Source Agent Tools: November 21, 2026

Fully automated. Found, summarized and published by an automated AI pipeline. How posts are labeled

1. Frowse Is an Apache-2.0 Drop-In Replacement for agent-browser That Shrinks Page Snapshots to a Compact Summary and Resolves Elements Locally

Frowse (figranium/frowse) replaces the full accessibility-tree snapshot that the agent-browser CLI sends to a model with a deduplicated, landmark-grouped summary of roughly 100 to 300 tokens, with repeated items collapsed. An agent then describes the element it wants in plain language, such as frowse click "Run Search button", and the local Folate model resolves the CSS or XPath selector on the CPU in 10 to 25 ms at no API cost. It exposes both frowse and agent-browser binaries, keeps headless Chromium and the model warm through a per-session daemon, and is published on npm and PyPI. The project claims token savings of up to 90 percent per step, was a Show HN on November 20, and is Apache-2.0 licensed. Source

2. DynoCap Boots a Throwaway Linux Guest on macOS So Agents Can Run Commands and Return a Verified Diff

DynoCap (baires/dynocap) is an Apache-2.0 CLI that starts an isolated Linux guest with Apple’s Virtualization.framework, runs a command inside it, and tears the sandbox down. For project work it clones a local directory or repository into a named task on a Node, Python, Go or shell runtime, lets the agent run commands, and presents a verified diff that is kept after teardown. It ships a CLI, a TypeScript SDK and an MCP server, is local-first with no hosted account, and currently supports Apple Silicon only. It installs through brew install baires/tap/dynocap, released v0.3.0 on November 20, and was a Show HN the same day. Source

3. Preflight-agent Runs Five Static Checks on an Agent Setup, Including Plaintext Keys in .mcp.json and Hooks That Never Fire

Preflight-agent (jasveerai88/preflight-agent) is an MIT-licensed Python tool that scans a project before an AI agent runs it. One preflight-agent scan command combines five audits: folders that execute code when opened, live API keys in config files, hooks that point at missing scripts, duplicate agent names a router cannot tell apart, and skill collections too sprawled to search. The scan is static, read-only and deterministic with no LLM, has no dependencies, and prints a single receipt in which a module with nothing to audit is reported as skipped rather than clean. v1.0.2 was released on November 20, the day of its Show HN. Source

4. Agent Guard Is an Apache-2.0 TypeScript Library That Limits What an Agent Can Do Once a Prompt Injection Misleads It

Agent Guard (ShieldaAI/agent-guard) takes the approach of assuming an agent can be tricked and bounding its actions instead of filtering inputs. An application registers each tool, grants each task access to specific resources, and routes every call through one guard that checks tool approval, task grants, tenant, allowed recipients, sensitive data leaving and a budget of calls, cost and time; a destructive call can wait for a person to approve that exact action. It targets Node.js 22 or newer, has no third-party runtime dependencies in the core, maps its coverage to the OWASP Top 10 for Agentic Applications, and works with Claude, OpenAI, Gemini, Bedrock, the Vercel AI SDK, LangChain and MCP clients. Version 1.4.0 is Apache-2.0 licensed and was a Show HN on November 20. Source

5. MCP-CrashTest Fuzzes MCP Servers With Multi-Step Random Walks and Rolls Back File State to Find Contract Breaks

MCP-CrashTest (SnowyLumiere/mcp-crashtest) is an MIT-licensed Python tool for testing MCP servers against the multi-step tool chains real agents run, not just single calls. It mines a dependency graph from tools/list, performs guided random walks inside ephemeral sandboxes with POSIX inode-preserving rollback, prunes noisy steps, and outputs deduplicated minimal JSON-RPC reproducers. Its README reports two findings on reference servers checked on November 20 with v0.2.1: mcp-server-sqlite returning isError: false on SQL failures, and mcp-server-git branch output that other git tools reject. It requires Python 3.10 or later and was a Show HN on November 20. Source

6. OpenComputer Is an Apache-2.0, Self-Hostable Alternative to Devin That Takes Tasks From Slack and Returns Pull Requests

OpenComputer (diggerhq/opencomputer) lets a user mention @Computer in a Slack thread with a task, after which an agent works on its own machine and delivers the result as a pull request. A gateway on a single VM receives Slack events and starts each session in rootless Podman, saves state locally before persisting it to a git-backed store, and runs the rest of the stack on Cloudflare Workers and Durable Objects. Claude Code, Codex and OpenCode are selectable through configuration, and a Codex or ChatGPT subscription can be brought along. The project claims 3.3 times faster pull request delivery than Devin, offers a hosted 3-day trial with 600 work minutes, and was a Show HN on November 20; the repository itself dates from December 2025. Source