🛠️ The Org Chart Your Agents Just Got

Repopaperclip

100,970 stars · paperclipai/paperclip

An open-source orchestration layer that manages a team of AI agents the way you'd manage a company, not a pipeline.

Here is the real problem it solves. Once you're running more than one agent, the hard part stops being prompts and starts being coordination. Who's doing what, what it costs, whether it's actually aligned to the goal you set. Paperclip gives every agent a role, a budget, and a reporting line, and gives you one dashboard to watch all of it.

Quick start: clone the repo, define a goal in plain language, then assign agents to it from any harness you already run (OpenClaw, Claude Code, Codex, Cursor, Gemini CLI, and others plug straight in).

Catch: this is infrastructure for a team of agents, not a single assistant. If you're running one agent on one task, you don't need a company, you need a to-do list.

Open the repo

AffiliateHubSpot

HubSpot, a CRM where AI agents work on the customer record your team already keeps. A Customer Agent answers support, a Prospecting Agent researches and drafts outreach, a Data Agent cleans what sits underneath, and you can build your own on one canvas from your prompts and CRM data. The CRM underneath has a free tier.

Use it for: putting support, prospecting and data-cleanup agents on the record your team already keeps.

See how HubSpot works

We may earn a commission.

Repowshobson/agents

40,509 stars · wshobson/agents

A multi-harness plugin marketplace: one set of agents, skills, and commands that runs across Claude Code, Codex, Cursor, OpenCode, Copilot, and more.

Use it for: stop rebuilding the same code-review or testing skill for every harness your team happens to use. 202 agents and 184 skills, installed once, portable everywhere.

Catch: the catalog is big enough that picking the right plugin takes real browsing. Budget ten minutes before you expect to be productive.

ToolOpenHands Agent Canvas

90,750 stars · OpenHands/OpenHands

A self-hosted control center that runs OpenHands, Claude Code, Codex, Gemini, Pi, OpenCode, or any ACP-compatible agent across local, remote, and cloud backends.

Use it for: if Paperclip is the org chart, this is the machine room, one place to start conversations and automate recurring jobs like turning a GitHub issue into a task queue.

Catch: runs locally by default, but the useful stuff (Docker containers, VMs, cloud backends) takes setup before it pays off.

Repowheresryan22/anatomy

6,253 stars · wheresryan22/anatomy

A Claude skill that explains a technical idea by building it as an interactive isometric illustration instead of a paragraph.

Use it for: ask it to explain a rate limiter or a cache key, and it hands back a working animated diagram instead of prose. Runs in OpenCode and Codex too, no API key required.

Catch: it's a presentation layer, not a debugging tool. Great for explaining a system to a teammate, not for finding the bug in it.

You are the person who has to clear up afterwards when something automatic goes wrong on a Friday afternoon. Sort my list of repeated jobs with that in mind.

The jobs I do over and over, with roughly how long each takes and how often: [THE JOBS YOU DO AGAIN AND AGAIN, ONE PER LINE, WITH ROUGHLY HOW LONG EACH TAKES AND HOW OFTEN]
What already goes wrong when one of them is done badly: [WHAT ALREADY HAPPENS WHEN ONE IS DONE BADLY]
Who would notice a quiet mistake, and how long that would take: [WHO WOULD SPOT A QUIET MISTAKE, AND HOW LONG THAT WOULD TAKE]

Give me, in this order:

1. The payoff, job by job. How long it takes, multiplied by how often, using only the figures I gave you, with the arithmetic shown on the line. Where I gave no time, write time not supplied. Where I gave no frequency, write how often not supplied. Leave the payoff off those lines rather than estimating it.

2. Four questions per job, answered short. Does it come out the same way every time, or does it change. What would it cost me if it went wrong and nobody spotted it for a week. How often does something odd turn up that a person has to stop and think about. Does anybody outside the business see the result.

3. The four piles, with every job I listed in exactly one of them.
Hand it over now: worth the time, the same every time, and a mistake would get spotted.
Hand it over with somebody checking: worth the time, but a quiet mistake is expensive, so a person looks before it leaves.
Fix the job first: the job itself is a mess, and doing a mess faster is worse than doing it slowly.
Keep doing it yourself: not worth the time it saves, or the judgement is the whole job.

4. The one I will want to automate and should not, named out of my own list, with the reason in a sentence I would repeat to somebody else.

5. What of this came from general knowledge. Anything you said about how a job of this kind usually behaves, rather than from what I told you, listed and marked as general knowledge for me to check.

Rules. Work only from the jobs I listed. Do not name a tool, an app or a service; this is about which jobs, not about what to buy. Do not invent a time or a frequency I did not give you, and do not tell me how many hours a week I would save unless every figure in that sum came from me. If I wrote "nothing goes wrong", ask me what a bad one would look like rather than treating the job as safe, and say which piles that answer would move. If I wrote "nobody would notice", no job on my list goes in hand it over now; the best it can do is hand it over with somebody checking, and say why. Before you sort anything, if a job on my list happens fewer than a handful of times a year, say so and put it straight in keep doing it yourself.