ivo's home

codex-keysmith is a Python 3.10+ command-line utility that manages instruction injection for local Codex instances on macOS and Linux. It runs via a single signed release script or git checkout, letting you preview configuration plans with dry-run flags before writing to disk. The tool modifies only downstream session behavior without reading account credentials or altering the core client binary.

If you manage multiple local AI coding agents and want reproducible, auditable system prompts without manual file edits, it is worth a look. However, Windows support remains experimental and unverified, and macOS users will trip over Gatekeeper due to unsigned binaries. Skip it if you prefer managing environment variables and configuration files by hand.

more: https://github.com/Jia-Ethan/codex-keysmith

Octop is an open-source, self-hosted Python application that runs a multi-agent AI assistant platform locally. A single process handles the FastAPI backend, SQLite database, React dashboard, and integrations with Telegram, DingTalk, and other messaging apps. It supports RAG, browser automation via headless Chromium, and IDE workflows through the agent-client-protocol.

It is worth setting up if you want to centralize private chat logs and scheduled tasks on your own hardware without relying on third-party SaaS wrappers. However, managing the configuration, security guardrails, and multiple agent profiles requires patience and a tolerance for early-stage Python tooling. Skip it if you just want a standard single-user chat interface.

more: https://github.com/TencentCloud/Octop

AIPOCH Open-Science is a local-first, model-agnostic desktop application for running reproducible research tasks via AI agents. Released at version 0.33.2 with cross-platform support for macOS, Windows, and Linux, it coordinates Python and R execution, literature search, data connectors, and file generation within a single project workspace. It functions essentially as an orchestrator for local or API-backed LLMs to automate data-intensive analysis while maintaining traceable provenance.

If you routinely juggle scripts, data tables, and literature notes across different notebooks, this workbench saves significant friction by keeping the execution loop contained locally. However, it is not built for teams wanting collaborative cloud whiteboards or anyone unwilling to vet agentic code output before it runs on their filesystem. Expect to spend real time configuring your local runtimes and API keys during the initial setup wizard.

more: https://github.com/aipoch/open-science

GenOffice is a cross-platform desktop application and CLI that edits native docx, xlsx, pptx, and pdf files using local or cloud AI models. It preserves the underlying file structures byte-for-byte while allowing an integrated agent to write formulas, generate slides, and modify documents with tracked changes. The project includes a local SQLite text search and supports custom API keys for major providers alongside a free tier.

It is worth testing if you routinely script document generation or want an agent that can interact directly with office formats on your local machine. The catch is that while the file handling is local, the heavy lifting depends on external LLM calls and API costs can add up quickly. If you do not use coding agents or AI assistance in your document workflow, standard office suites remain simpler.

more: https://github.com/genspark-ai/genoffice

DSCODE is a terminal-based coding agent for macOS built on DeepSeek Harness. It uses a persistent shell to run tests and edit files while allowing multiple local sessions to share the same runtime and communicate directly. You can script non-interactive turns, run background side-tasks in read-only child sessions, and delegate work between agents using a shared mailbox.

It is worth a look if you already rely on DeepSeek and want multi-agent workflows running locally in your terminal rather than in a cloud IDE. However, it is strictly tied to macOS and requires trusting its automated review model to handle permissions and diff checks. Skip it if you use Linux or prefer a traditional single-agent setup.

more: https://github.com/qiz029/dscode

This is a plugin for Codex and Claude Code that enforces strict boundaries on AI task execution. It intercepts tool calls to prevent unauthorized file writes, unnecessary dependency additions, and redundant verification loops. The system runs on Node 18+ and uses a specific skill and guard architecture to halt actions that violate the defined intent.

It is useful if you are tired of agents refactoring code they were not asked to touch or generating useless checksums. The catch is that it requires manual configuration for specific file or dependency allowances, which adds friction to complex tasks. This is not for teams that prefer fully autonomous agent behavior without hard stops.

more: https://github.com/lennney/stop-that-shit