AI Agent Repos and Starter Templates: Seven Projects
Seven open-source projects for building AI agents, from a small starter template to full libraries, with what each is for and how to make one your own.
These are seven open-source projects for building AI agents, from a small starter template to the libraries larger agent apps are built on. If you're new, start with the ReAct Agent template, because it's small enough to follow every step the agent takes.
This isn't a ranking by GitHub stars, and none of these is a finished app you can ship as is. To try three of them hands-on and pin one, work through the agent starter lab from the course.
Last verified Sep 17, 2026. Recheck target: monthly. The CSV holds all 7 entries.
Start with five
Pick the starter template if you want to see how an agent works end to end. Pick a library if you already know your hard part: picking a long task back up where it left off, getting answers back in a fixed format, coordinating tools, or running a coding agent. Add a chat screen only when you already have a backend that needs one.
A small app where you can see model calls, tool calls and how the state changes, all in a few files. The companion starter lesson locks to a specific commit of this repo and runs its real functions with a fake model.
Pick this when your workflow needs to save its progress, branch into different paths and pause for a human to review. Read the examples first, because a simple loop doesn't need a graph.
Good when the rest of your app needs the agent's output checked and its dependencies typed, so nothing arrives in the wrong shape. Start with one tool and a defined output type before adding anything bigger.
The examples show tools, handoffs between agents and tracing working together. Read them to see how the pieces fit before you build a bigger workflow on top.
Start here when your app needs Claude's agent tools and permission controls. Check which tools are allowed and which hooks run before you point it at a real workspace.
The directory
Each link goes to the maintainer's own repository. Before you use one, read its README, examples, license, release notes and security guidance for the version you plan to install. Being on this list doesn't mean a project passed a security review or that its demo is ready for production.
Pick this when your workflow needs to save its progress, branch into different paths and pause for a human to review. Read the examples first, because a simple loop doesn't need a graph.
README inspected at commit 448a76377956; the repository was not archived when checked. The activity date is the default branch commit timestamp, not a release date or a reproduced integration test. This monorepo has separately versioned packages; a client SDK release is not the graph library version.
Good when the rest of your app needs the agent's output checked and its dependencies typed, so nothing arrives in the wrong shape. Start with one tool and a defined output type before adding anything bigger.
README inspected at commit 532fb9b98092; the repository was not archived when checked. The activity date is the default branch commit timestamp, not a release date or a reproduced integration test.
A small codebase that shows how a model writes code and then runs it. Where and how that code runs is part of the design, so read the execution and sandbox choices closely.
README inspected at commit 30bb1161095d; the repository was not archived when checked. The activity date is the default branch commit timestamp, not a release date or a reproduced integration test.
The examples show tools, handoffs between agents and tracing working together. Read them to see how the pieces fit before you build a bigger workflow on top.
README inspected at commit e34311b973cb; the repository was not archived when checked. The activity date is the default branch commit timestamp, not a release date or a reproduced integration test.
Start here when your app needs Claude's agent tools and permission controls. Check which tools are allowed and which hooks run before you point it at a real workspace.
README inspected at commit 9d398b6b7ed8; the repository was not archived when checked. The activity date is the default branch commit timestamp, not a release date or a reproduced integration test.
A small app where you can see model calls, tool calls and how the state changes, all in a few files. The companion starter lesson locks to a specific commit of this repo and runs its real functions with a fake model.
README inspected at commit 9bbd82d84905; the repository was not archived when checked. The activity date is the default branch commit timestamp, not a release date or a reproduced integration test.
Use this if you already have a LangGraph backend and want a chat interface on top. It won't check who your users are or what they're allowed to do, so you still have to build that.
README inspected at commit 41926d89c979; the repository was not archived when checked. The activity date is the default branch commit timestamp, not a release date or a reproduced integration test.
What each one is for
| What you need | Try | What you still have to supply |
|---|---|---|
| See how a small agent makes decisions | ReAct Agent template | Your task, test cases, tools and launch controls |
| Pick a long workflow back up where it stopped | LangGraph | What gets saved, save points, permissions and a recovery plan |
| Get Python output in a fixed, checked format | Pydantic AI | Your business rules, error cases and storage |
| Study agents that write and run code | smolagents | A safe place to run that code, and limits on it |
| Build around tools, handoffs and tracing | OpenAI Agents SDK | Your task logic, tool permissions and tests |
| Run a coding agent inside your app | Claude Agent SDK | Which files it can touch, runtime compatibility and who may use it |
| Put a chat screen on a LangGraph backend | Agent Chat UI | A backend, plus login and permission checks |
What was checked: each repository's identity, its pinned docs and its recent activity. None of the seven was deployed or compared on a live model for this page.
For a first look at the ReAct template, find three places in the code: where it calls the model, how it picks a tool, and what sends the workflow on to the next step. Then compare those with your own task.
Make a starter your own
Before you connect real customer data, write down three things: the task, the tools the agent may use, and when it should stop. Run one official example with throwaway inputs. Then swap its example prompt for a small set of test cases from your real task. Include one case that should fail and one request the agent should refuse.
Pin your dependency versions and keep the example's license notices. Add timeouts, a limit on cost or turns, and a log of every model and tool call. If a tool can change something outside your machine, require a person to approve it, or point it at a test environment you can reset. A good run leaves something you can inspect, not just a confident final message.
When you compare frameworks, give each one the same task and inputs. Write down whether it got the right answer, how it recovered from a failure, and how hard its trace was to follow. More features only help if your app needs them.
Separate the framework from where it runs
A Python library can coordinate an agent without giving it a separate, safe machine to run untrusted code. A hosted runtime can give you that machine without deciding how the agent thinks. The agent development environments guide separates those choices.
For customer-facing engineering jobs, compare what each role owns in the FDE and GTM hiring directory. For the methods underneath, follow the AI engineering blogs. For a reading routine you can import, use the RSS pack.
Change log
Expanded from five frameworks to seven projects, including an agent application template and a chat frontend. Pinned README and default-branch activity checks are recorded for every entry.
Added a focused selection with primary-source links, a start-with-five and an export.
Follow changes to every directory with the directory changes RSS feed.