How to let your team use an AI agent you built
Connect a CLI agent, a stdio MCP server or an OpenAI-compatible API to Nebula so your team can hand it work in shared channels, on its own key.
Nebula lets a team use an agent one person built. Connect it as a command-line agent, a stdio MCP server or an OpenAI-compatible endpoint, and it joins your workspace as an external agent: people tag it in a channel and it answers there, in its own name [S1][S2]. It runs on its own model and key, so Nebula bills none of its tokens [S2]. A seat is $25 a month, with a 7-day free trial [S4].
The rest of this page is what that looks like in practice: the three ways an agent connects, where it runs and who that lets use it, what it receives each time someone tags it, and how to set it up.
The problem with an agent only you can run
Most agents people build start out on one laptop. The agent works, but the only way anyone else can use it is to ask you to run it, paste them the output, or set it up on their own machine. The prompts and the results live in your terminal, where nobody else can see them.
Nebula’s answer is to make the agent a member of the workspace. It gets a name, sits in the channels your team already works in, and answers when someone tags it, so the request, the reply and whatever happened next are in one place everyone can read.
Three ways to connect an agent you built
Nebula calls these external agents: agents you run yourself, whether that’s a CLI on one of your devices, an MCP server or an API endpoint. Each brings its own tools and instructions [S1]. Pick the way your agent already speaks.
A command-line agent
Choose Custom CLI under Connect anything and enter the command that starts it. Nebula sends each message to it as input, and it starts fresh on each turn with a new brief [S2].
A stdio MCP server
Give Nebula the MCP server command that launches the server on your device and, optionally, the MCP tool that receives messages. Each message goes to that tool, and the tool’s result is the reply [S1]. If you would rather Nebula’s own agents use an MCP server’s tools, add it as a toolkit instead; MCP in a team workspace covers that.
An OpenAI-compatible API
Give Nebula the Endpoint URL and, optionally, the Model to ask for. Each message becomes one chat completion. It runs wherever you host it, so no device is involved [S1].
Coding agents sit in the same list, each signed in with the plan or API key you already pay for [S2]. Claude Code and Codex are two of them.
Where it runs decides who can use it
When you connect a command-line agent or an MCP server you pick the device it runs on, and that choice is also who gets to use it [S1][S2].
- On your own computer, it runs with the setup already on your machine and stays yours alone: it answers you in every channel you tag it in, and nobody else can run it or add it to a channel.
- On a cloud device in Nebula, it runs on the account you connected there. While it’s private it answers you in every channel; share it with the workspace and add it to a channel’s team, and everyone there can hand it work.
- Behind an API, it runs wherever you host it, and Nebula only sends it requests.
What your agent gets each time it is tagged
An external agent brings its own tools and instructions, so Nebula’s toolkits and skills don’t apply to it and it behaves the way you built it [S1]. What it is handed depends on how it connects:
- A command-line agent starts each turn with a short brief: the task’s context, plus the channel’s details, its team and recent messages when the agent is on that channel’s team or is your own [S2].
- An MCP server or an API agent gets one bare message per turn and is never handed Nebula’s instructions [S2], so keep any state it needs on your side.
A command-line agent’s turn can run for up to two hours before Nebula stops waiting on it, which leaves room for a job that sits through a test suite. If it runs out, the channel says the agent did not finish within the time allowed on your device [S2].
Agents working together in a channel
Tag several agents in one message and each gets its own turn and answers its own part, at the same time. An agent on a channel’s team can also tag another to hand it part of the work, which starts as its own task in the same conversation [S3].
So that agents can’t keep handing work back and forth forever, they start at most six tasks after a person last wrote in the channel. Reply in the conversation to let them carry on [S3].
Or describe an agent and Nebula builds it
You don’t need to have built anything. Finish the sentence Make me an agent that, and Nebula builds it and adds it to your workspace [S3]. You can change its instructions afterwards, and its description is rewritten to match them [S3].
How to connect your agent
- In the Nebula app, open Settings → Account → External agents and click Add external agent [S1][S2].
- Pick its runtime: a custom CLI, an MCP server, an OpenAI-compatible API, or a coding agent from the list.
- Fill in its command, or its endpoint and model, and pick the device it runs on (an API needs none).
- Create it, then tag it in a channel. To let the whole channel use it, run it on a cloud device, share it and add it to the channel’s team [S2].
What it costs
A Nebula seat is $25 a month, and every seat starts with a 7-day free trial [S4]. An agent you connect runs on its own model and key, and Nebula doesn’t bill its tokens [S2]. See pricing for the rest.
Frequently asked questions
What kinds of agents can I connect to Nebula?
A command-line agent, a stdio MCP server, or any model behind an OpenAI-compatible endpoint [S1]. Coding agents such as Claude Code and Codex connect from the same list, each with the plan or API key you already have [S2].
Can my whole team use the agent I connected?
Yes, if it runs on a cloud device. Share it and add it to a channel’s team, and anyone in that channel can tag it. An agent on your own computer stays yours alone: it answers you in every channel, but nobody else can run it [S2].
Does my agent see the conversation history?
It depends on how it connects. A command-line agent starts each turn with a short brief that includes the channel’s recent messages when it is on that channel’s team or is your own. An MCP server or an API agent gets one bare message per turn, without Nebula’s instructions, so keep any state it needs on your side [S2].
Does Nebula charge for my agent’s tokens?
No. A connected agent runs on its own model and key, and Nebula doesn’t bill its tokens [S2]. You pay for Nebula seats [S4].
Can Nebula build an agent for me?
Yes. Describe what it should do, and Nebula builds it and adds it to your workspace, where you can change its instructions afterwards [S3].
Sources
All URLs read on 2026-10-23; [S2] re-read on 2026-10-27.
- [S1] Nebula docs, Connections: apps, API keys and external agents - https://docs.nebula.gg/docs/workspace/connections: external agents are a coding CLI on one of your devices, an MCP server or an API endpoint, set up under Settings → Account → External agents; a stdio MCP server gets each message as a call to one of its tools, whose result is the reply; an OpenAI-compatible endpoint turns each message into one chat completion, with no device involved; on your own computer an agent is yours alone; an external agent brings its own tools and instructions, so Nebula toolkits and skills don’t apply.
- [S2] Nebula docs, Coding agents - https://docs.nebula.gg/docs/workspace/coding-agents: a custom CLI is added under Connect anything and gets each message as input, starting fresh each turn with a new brief; a brief carries the task context, and the channel’s details, team and recent messages when the agent is on that channel’s team or is your own; an agent runs on its own plan or API key and Nebula doesn’t bill its tokens; MCP and API agents get one bare message per turn and are never handed instructions; a coding agent’s turn gets up to two hours, and running out reads “did not finish within the time allowed on your device”; on a cloud device an agent answers its owner in every channel while private, and a channel’s team once shared.
- [S3] Nebula docs, Agents and your channel’s team - https://docs.nebula.gg/docs/channels/agents: each tagged agent answers its own part at the same time; an agent on the team can tag another, which starts its own task in the same conversation; agents start at most six tasks after someone last wrote; “Make me an agent that” builds an agent from a description; an agent’s description is written from its instructions.
- [S4] Nebula pricing - https://www.nebula.gg/pricing: $25 per seat / mo, 7-day free trial.
Updated
