Switchroom
A switchboard for your Pro or Max. Your Claude subscription, as a fleet of always-on specialist agents you talk to from Telegram. Opinionated UX, done properly.
I loved OpenClaw + Telegram. I wanted my Claude subscription. And the UX done properly. So I built this.
Latest release notes →
See what your agent is doing
Every time an agent starts work, a progress card pins into its Telegram topic and updates in place as tools execute. Each Read, Bash, Edit, Grep is visible as it happens, with elapsed time so you can tell if something's stuck. Sub-agents surface in the same card. When the agent finishes, the card flips to Done and unpins.
No silent gaps. No ghosts. No squinting into a black box.
⚙️ Working… · ⏱ 12s
💬 refactor the auth module to use JWT
─ ─ ─
… (+3 more earlier steps)
✅ Read src/auth/session.ts
✅ Grep "cookie" (in src/)
🤖 Edit src/auth/jwt.ts · 4s
The card is the headline UX. The rest of the product is in service of it.
- Cards update at most once every 5 seconds. Fast enough to follow, not so fast it floods.
- Last 5 steps stay visible. Older ones collapse into
(+N more earlier steps).
- Running steps show elapsed time so a stuck tool is obvious.
- Tool labels are deterministic, written by a
PreToolUse hook, so the card never lies about what's running.
- Two agents working at once? Each gets its own card, labelled
(1/2) and (2/2).
Sub-agent visibility
When an agent delegates to a sub-agent — Opus plans, Sonnet implements — the sub-agent's work shows up indented inside the parent's pinned card. One pinned surface per task, however many processes it spawns underneath. Nothing gets buried in a side-channel you have to go look for.
Right, so what's this about
So you had the bright idea. Run Claude Code agents 24/7 on a cheap Linux box, talk to them from Telegram, use the Claude Pro or Max subscription you're already paying for. Sensible. Obvious, even.
Then you tried OpenClaw. Followed the docs, spun it up, got it running, only to realise halfway through that you're pinging the Anthropic API on your own key and your token bill is quietly ticking over in the background. Bit of a bait and switch, that one. You signed up for "use your subscription," not "buy API credits on top of your subscription."
So you gave Claude Code's built-in Telegram channel a crack instead. Sent a message. Waited. Something happened, maybe. Eventually a reply came back. What did the agent actually do? No idea. Which tools ran? No idea. Did it get stuck, crash, spawn a sub-agent, read half your repo? No idea. It's an MVP black box of death, and I got sick of squinting into it.
So I built this.
What you get
