One room for terminals, coordinated agent work, review, Design, and a partner whenever the build needs another mind.
The room for directing coding agents. Connect a bridge on the machine that holds the project, then run Claude Code, Codex, or Hermes Agent across visible terminals. Split work into coordinated lanes, review what comes back, and bring working interfaces into Design. Start alone or invite a partner into the same live room.
A single running instance of a coding agent inside your room. Open more than one to work on different things at once or look straight into what another terminal is doing. Every terminal stays visible in the room. Top-level agents with room tools can inspect other terminals and delegate a defined slice to a visible worker lane.
The room still opens and every past terminal is still readable, but nothing new happens. A room with no bridge connected is a read-only transcript, not a chat: there is no fallback conversation and nothing answers on its behalf.
thinkpool supports Claude Code, Codex, and Hermes Agent. Claude Code uses the host’s normal Anthropic login by default, or a registered Anthropic-compatible provider for that lane. Codex uses the host’s Codex/OpenAI login; custom bridge providers are not wired to Codex. Hermes Agent uses an isolated thinkpool ACP profile, set up before its first lane. The bridge does not include an agent: install a supported runtime on the bridge machine and give it the login or credentials it requires.
Yes. You can run multiple agent terminals in the same session and choose either Claude Code, Codex, or Hermes Agent for each one. The selected runtimes must be installed and configured on the bridge machine.
Free Dispatch runs up to 3 active worker lanes per room; Plus runs up to 6. Parent terminals do not use that worker allowance. Every parent and worker terminal shares a 12-live-lane safety ceiling on the bridge machine. An agent can open at most six new lanes in a rolling ten-minute burst, and a worker cannot open another worker. Human-opened parent tabs do not use the burst allowance, but they still count toward the host ceiling. New Dispatch tasks queue while the host is full or under memory pressure, then start when capacity returns.
Yes. Register an Anthropic-compatible endpoint in the bridge menu and it becomes available in the model list for Claude Code terminals. Locally hosted LLMs can also be connected through the same kind of endpoint.
We can never read a custom provider key. Adding one seals it in your browser before it is sent to the bridge. The sealed envelope passes through our servers on its way to your host machine, which is the only thing holding the key that opens it. That private key is generated on your machine and never leaves it. The encrypted form crosses our infrastructure; the readable key never does, and we cannot recover it, store it, or log it.
Free includes one active hosted room and up to 3 active dispatched worker lanes per room. Plus has no hosted-room count cap and raises Dispatch to 6 active worker lanes per room. Every parent and worker terminal shares a 12-live-lane safety ceiling on the bridge machine. New Dispatch work queues while the host is full or under memory pressure, then starts when capacity returns. Close a hosted room and you can start another straight away; there is no monthly allowance to run out of. Plus costs €5/mo flat, with no trial clock, annual plan, or student rate. Both plans allow up to 4 files per message, 25 MB each.
They join with their own free account through your room link, and can see every terminal and steer alongside you.
Session and terminal history lives on Supabase in the EU (Frankfurt, Germany). The repository stays on the bridge machine; files deliberately attached to a room and agent-published images or previews are stored separately in private Supabase Storage. A custom Claude provider key is sealed in the browser and relayed as an encrypted envelope that only the bridge can open. See How your data moves for the full boundary and retention details.