Top Picks

A detailed post from pact0 about testing whether blank AI agents could complete a real product workflow, what failed, and the fixes they shipped after the audit.

ObserverBench is an interactive benchmark and guided demo for testing whether internal AI monitoring methods help make safer decisions, not just better predictions. It includes real-model results, benchmark tasks, and a no-login walkthrough showing how warning scores affect review and harm that slips through.

OpenAI’s documentation for its Agents API, covering how to build and run cloud agents with the OpenAI platform. The page surfaces core concepts, quickstart material, SDKs, webhooks, tracing, tools, and related agent-building resources.

A90 presents a tiny guestbook canvas where people draw and sign with passkeys, while AI agents can submit artwork through a documented proof-of-work flow. The page explains the human and agent paths, the submission protocol, and how the Imprint Index works.

LeanGuard is a Lean 4-based runtime guardrail system for agent tool use, with native temporal policies, Python integration, and optional MCP support.

A web calculator that estimates the electricity use and carbon emissions of AI token usage, with everyday comparisons, example workloads, and a detailed methodology page.

Open-source text-to-SQL engine tuned to your own database, with configurable pipeline stages, policy checks, and an evaluation harness for measuring performance on real schemas.

A Linux desktop app for managing multiple AI helpers with separate roles, memory, provider connections, and isolated computer workspaces.

Hazzel is a small terminal coding agent for working directly in a codebase. It reads files, edits with diff approval, runs commands with permission, and supports git-aware actions, with multiple model providers and bring-your-own-key setup.

Open-source ticket system release with internal agent chat, ticket handovers, scheduled report delivery, customer/web form processes, CMDB links, and FAQ attachments.

brw is a semantic browser control tool for agents. It runs a real visible Chrome/Chromium browser and exposes control through MCP and HTTP, with stable refs, semantic page snapshots, and post-action observations.

Self-hosted Linux monitoring agent that reports system health and security events through a private Telegram bot. It covers CPU, memory, disks, temperatures, battery, services, SSH logins, brute-force detection, and remediation actions.

Local-first SQLite memory engine for AI coding tools that preserves project decisions and context across agents and model switches.

PicEditor is a browser-based AI photo editor and image generator that lets users edit uploads with text prompts or create new images from text, with no sign-up required. It supports background replacement, object removal, text edits, style changes, and multiple model options.

A fully local voice-controlled 3D sandbox where a sassy robot turns spoken commands into voxel builds. The repo explains the stack, runtime requirements, service contracts, controls, and tests.

An open-source Stripe-like payment layer for AI agents and paid MCP tools, with spend authorization, per-call charging, and settlement flows for test credits in beta.

A playable Three.js office for local Herdr agents, with live terminals, coffee breaks, and a shotgun. It includes run instructions, configuration flags, controls, and notes on live vs demo mode.

OpenSpender is a self-custodial wallet for agent-to-agent payments. The page shows a command-line connect flow, explains allowance-based spending rules, and includes an example of one agent paying another through MCP-compatible tools.

A self-hostable MCP endpoint that centralizes accounts, memory, skills, secrets, and access controls for AI agents. It supports multiple providers, profile-based permissions, audit logs, and deployment on your own machine or cloud.

An open-source collection of Claude Code skills, agents, and plugins for improving agent output quality across tools like Cursor, Gemini CLI, OpenCode, and Antigravity.

A local CLI that records a codebase’s behavior and gates AI-driven refactors against a frozen baseline. It runs declared probes, compares outputs byte-for-byte, and returns a simple pass/fail signal for agent workflows.

Bridle is an open protocol and small CLI for handing context and work between coding agents on different machines, with opt-in policy controls on both sides.

WorldCut is a TypeScript package that checks whether observations from independent systems satisfy declared version and time relationships before an autonomous agent makes a decision.

Subpool is a self-hosted control plane for managing AI subscription and API capacity across teams. It pools authorized accounts, issues employee-specific keys, exposes OpenAI-compatible endpoints, and tracks usage without storing prompts or responses.