Records we built ourselves and open tools we point to. Read what each one is, and what to check before adopting it. We do not rank or benchmark open source — only what it is for and what to check.
Last checked 2026-10-17 · rechecked monthly
Our own write-ups, public assets, and videos.
A development log series: one release note per Marblo release, plus a deep-dive post per feature.
For enterprise adoption
Shows where approval gates sit when several agents run in parallel, as actual design decisions. Start from the release note and follow links to the feature posts.
A hands-on comparison of how Codex, Grok, and Claude Code resume sessions, with the versions tested.
For enterprise adoption
If you plan to mix several coding agents in one workflow, this shows where a failed resume ends the process instead of warning.
Why judging agent activity from terminal output misleads in both directions, and how to define state by boundaries instead.
For enterprise adoption
Useful when deciding how to tell "stalled" from "working" for long-running agents in operations.
A diagnosis of the wiring point that fails silently when you add a new harness to per-agent cost tracking.
For enterprise adoption
For teams splitting AI spend by department or agent: a checklist order for when the cost table comes back empty.
A registry of skills, agent definitions, MCP servers, knowledge packs, and workflows published as plain files.
For enterprise adoption
Each item carries a grade (official, verified, community); read the grade definitions before bringing anything in-house. Community items are unreviewed.
What Marblo sends off the machine, which settings turn it off, and answers for a security review.
For enterprise adoption
Meant to be handed to a security team as a link. It also lists what is not yet true.
The Hypemarc founder's YouTube channel (Korean), covering AI agents and data work.
For enterprise adoption
For seeing a workflow on screen before reading about it.
A video (Korean) of agents building a small game end to end with the Marblo closed loop.
For enterprise adoption
Pair it with the closed-loop build note to see the written steps on screen.
Repository addresses and licenses were rechecked on the date shown. Archived repositories and repositories whose license could not be confirmed are left out.
'Conditional' marks repositories where the license differs by directory or file, or where extra conditions apply to offering it as a service or to branding. Items without the label use a single permissive license such as MIT or Apache-2.0 across the repository. This page is not legal advice; read the LICENSE file before adopting.
A framework for building stateful agent flows as graphs.
For enterprise adoption
For writing workflows with approval steps and retries explicitly in code. The flow stays auditable, but there is a learning curve, so keep the pilot narrow.
A framework for setting up several role-based agents that work together.
For enterprise adoption
Fits prototypes of role-split work such as research, drafting, and review. Set call limits and logging first; cost and tracing get harder as agents multiply.
A programming framework for agents that work by conversing with one another.
For enterprise adoption
For research and prototypes of multi-agent conversation. Check the repository notes for how it relates to Microsoft Agent Framework.
A framework for building and deploying agents and multi-agent workflows (Python and .NET).
For enterprise adoption
A candidate for organizations running both Python and .NET. It is relatively new, so pin versions and track the changelog.
A lightweight multi-agent framework built around tool calls and hand-offs between agents.
For enterprise adoption
A small starting point for tool calls and hand-offs. If you may switch model providers, check how far support for other models goes.
A Python library for agents whose output is validated against typed schemas.
For enterprise adoption
Fits work where agent output feeds other systems directly. Schema validation guarantees shape, not correctness, so evaluate content separately.
A minimal library where agents act by writing code.
For enterprise adoption
Small enough for teaching the concepts. Because it executes code, connecting it to production data requires an isolated sandbox.
An LLM application framework centered on document processing and data connectors.
For enterprise adoption
For connecting internal documents to search and Q&A (RAG). Document-level permissions must be designed into indexing separately.
An orchestration framework that assembles retrieval and agent pipelines from components.
For enterprise adoption
Components are easy to swap and inspect. Turn on per-component logging so you can see which documents fed each answer.
A TypeScript framework for AI applications and agents.
For enterprise adoption
A candidate when a web team wants to stay in TypeScript. Read the ee/ license separately if you plan to use those features.
A collection of reference implementations of Model Context Protocol servers.
For enterprise adoption
A baseline for learning how agents use tools such as files and search. Reference code may not cover production security, so narrow permissions yourself.
The official Python SDK for building MCP servers and clients.
For enterprise adoption
The starting point for exposing internal systems as MCP tools. Separate read and write per tool and log every call.
GitHub's official MCP server.
For enterprise adoption
For delegating issue and PR work to agents. Start with a read-only token and scope write access per repository.
An MCP server that exposes browser automation as tools.
For enterprise adoption
For web checks and repetitive form work. It hands a logged-in session to the agent, so separate accounts and restrict reachable domains.
A coding agent that runs in the terminal.
For enterprise adoption
For developer-team pilots. Agree on how agent-run commands are approved and how working directories are isolated.
An agent CLI for using Gemini models from the terminal.
For enterprise adoption
For running the same task side by side with other vendors’ CLIs. Include what code leaves for an external model in your security review.
An agent platform where AI carries out software development tasks.
For enterprise adoption
Fits experiments in delegating work issue by issue. Check in the install docs whether it runs in containers and how outbound network access is blocked.
An autonomous coding agent available as an IDE extension, CLI, and SDK.
For enterprise adoption
Works step by step inside the IDE with approvals. Align extension-install policy and API key handling with internal rules.
An AI pair-programming tool for the terminal.
For enterprise adoption
Changes land as git commits, which keeps review simple. As of the check date the last push was in May 2026, so look at recent issue and release activity first.
An extensible agent that goes beyond suggestions to install, run, edit, and test.
For enterprise adoption
For experiments extending an agent beyond development via extensions. The repository moved (formerly block/goose); this links to the current address.
An observability platform for tracing, evaluating, and improving LLM applications.
For enterprise adoption
Acts as an audit log of which input produced which answer. Logs will contain personal data, so decide masking and retention first.
A tool for observing and evaluating AI applications.
For enterprise adoption
For tracking experiments in-house. Get legal review before offering it to customers as a service.
An evaluation framework that checks LLM output repeatedly, the way tests do.
For enterprise adoption
For pre-release regression checks. Scores depend on configuration, so define and document the criteria yourselves.
An evaluation toolkit for LLM applications, often used for RAG answers.
For enterprise adoption
Good for learning how to check RAG answers against sources. As of the check date the last push was in February 2026 and the repository has moved, so confirm current activity.
A tool for testing prompts, agents, and RAG, including red-team checks.
For enterprise adoption
For finding weak prompts before release. It calls external models, so keep real personal data out of test sets.
A gateway that lets you call many LLM providers through one API format.
For enterprise adoption
A middle layer so switching providers does not mean rewriting the app. Every call passes through it, so key management, access control, and log retention are the core design.
A workflow automation platform with AI features.
For enterprise adoption
For automating repetitive work in non-engineering teams. Check the license before offering it to external customers.
A platform for building agent workflows and RAG in one workspace.
For enterprise adoption
Suits teams that want to prototype without code. Check the multi-tenant condition before any SaaS-style use.
A self-hosted AI chat interface for local models and external APIs.
For enterprise adoption
For standing up an internal chat UI quickly. Check the branding clause and how it integrates with your sign-in.
An inference and serving engine for LLMs.
For enterprise adoption
A candidate for serving models on your own GPUs to keep data in-house. Assumes GPU operations staff and budget.
A Python tool that converts Office documents and other files to Markdown.
For enterprise adoption
For cleaning internal documents before RAG ingestion. Sample-check for lost tables and formatting and keep a source comparison step.
A tool that converts PDFs and other documents into structures generative AI can use.
For enterprise adoption
A candidate for structurally complex documents. Test results on Korean-language and scanned documents yourself.
A crawler that collects web pages as LLM-ready Markdown.
For enterprise adoption
For automating research on public sources. Respect each site’s terms, robots policy, and copyright.
Vector storage and search infrastructure for AI applications.
For enterprise adoption
An easy store for small RAG prototypes. Design for production scale and access control separately.
A memory layer that gives agents persistent memory.
For enterprise adoption
For letting agents remember per-customer preferences and history. Memory is personal-data storage, so design a deletion path.
A Python library that lets AI agents operate a browser.
For enterprise adoption
For automation experiments on legacy web systems. The project also promotes a paid hosted service; read the open-source part separately.
A toolkit for adding programmable guardrails to LLM chat applications.
For enterprise adoption
For managing topic and format rules as code. Guardrails reduce risk but do not guarantee blocking, so pair them with separate testing.
A framework for detecting, masking, and anonymizing personal data in text and more.
For enterprise adoption
For masking personal data before it reaches an LLM. Korean ID and phone formats need custom rules you test yourself. The repository has moved; this links to the current address.
Follow the build notes and measured write-ups through RSS.
Reviewing this at work? Bring it to your team