Open Source Agent Tools: October 18, 2026

1. MIKE Runs Coding Agents From a Kanban Board Inside gVisor Sandboxes

MIKE. MIKE (MIKE: Integrated Kanban Environment), posted to Show HN on October 18 after its v0.1.0 release on October 15, is a self-hosted LLM harness where users define and delegate tasks as kanban cards and then review the agent’s deliverables. Each agent runs inside a gVisor sandbox with a git worktree as its workspace and a read-only repository mount for context, and extensions are isolated through iframe sandboxing and wazero. The interface is a Progressive Web App designed for low bandwidth and offline use, and it works with llama.cpp and any provider that supports the OpenAI Chat Completions API. It ships as a container image, uses passkey pairing for login, and is licensed under Apache 2.0, with no stable API or extension interface promised before 1.0. Source

2. Sphinx Makes Developers Pass a Quiz on Agent-Written Code Before Pushing

Sphinx. Sphinx, posted to Show HN on October 18, is a git hook that reads the diff of agent-written code before a push and asks the developer a question about how the code behaves when things go wrong. A correct answer lets the push continue, and a wrong one produces another scenario to work through. It works with Codex, Claude Code, Cursor, OpenCode, and Pi or with plain git, can use the existing coding agent, OpenRouter, or any OpenAI-compatible endpoint to generate questions, and never opens .env files or keys. A sphinx yolo command bypasses the check but logs every skip. It installs from npm as sphinx-understand and is MIT-licensed. Source

3. PrimerKit Gives Coding Agents a Workspace for Market Data and Paper Trading

PrimerKit. PrimerKit, released as v0.1.0 on October 17 and posted to Show HN the same day, is a workspace where an AI agent works with market data from an initial question through to a paper trade. The agent can brief users on recent market moves, answer questions while showing the data lines it used, backtest user-written rules against historical data, run those rules on a local schedule, and place simulated orders after approval, with every run saved to a local journal. It runs inside Claude Code, Codex, or Gemini CLI or in its own terminal, works with hosted models or local models through Ollama or LM Studio, and uses PreReason as its default data source. The project states the model never proposes strategies on its own and is licensed under Apache 2.0. Source

4. Busabase Offers a Shared, Reviewed System of Record for Multiple Agents

Busabase. Busabase, posted to Show HN on October 16, is an open-source database and workspace where Claude Code, Codex, Cursor, Gemini CLI, and other agents read and write the same structured records, docs, skills, and apps. Agent writes arrive as change requests with a field-level diff, author, and history, and workspace permissions decide whether each change merges immediately or waits for human review. Agents connect through an Agent Skill file, MCP, OpenAPI, a CLI, or an ACP chat session, and the server starts locally with npx busabase server on an embedded PGlite database, with Docker, desktop, and mobile clients also available. The project is MIT-licensed, and the same engine powers the hosted Busabase Cloud. Source

5. Termexo Lets One AI Operate Other Coding Agents’ Terminals on Windows

Termexo. Termexo, posted to Show HN on October 16 alongside its v0.10.14 release, is a Windows workbench that runs Claude Code, Codex, OpenCode, Pi Agent, Grok Build, and Antigravity in real terminals grouped by project. Its local MCP server lets a connected AI find another agent’s terminal, send and submit a prompt, and read the reply, and its task board can start tasks for five of the six supported agents. It also resumes native sessions, flags agents waiting for input or approval, and supports phone access through a self-hosted encrypted relay or a trusted LAN or VPN. The app is MIT-licensed and requires no Termexo account, with MCP control and remote access off by default. Source