The AI Terminal

Built not to confine your agents to a chatbox,
Coldtea is the best way to code in the terminal

Trusted by engineers from

Revolut

HelloFresh

Amazon

SuperTokens

Appsmith

LeoVegas

For those who think in Agent Teams

Pair Codex with Claude, or run a team of Claudes transparently. They talk to each other, and review work done, while you delegate

AGENT-TEAMS — ACTIVE
DarjeelingClaudeDriver
Claude Codev2.1.160Opus 4.8 (1M context) with max effort · Claude Max~/dev/coldtea
>add cursor pagination to /api/sessions
Write(api/sessions.ts)
+24
Edit(db/sessions.ts)
+8−3
Good catch — switching to keyset.
Edit(db/sessions.ts)
+2−1
>OFFSET paging table-scans on deep pages — switch to keyset on (created_at, id) before this ships.
>Press up to edit queued messages
SenchaCodexReviewer
>_OpenAI Codex(v0.136.0)model:gpt-5.5 xhigh/modelto changedirectory:~/dev/coldteapermissions:YOLO mode
Tip: Try the Codex App — run codex app
/review current changes
coldtea.pane_read
142 lines · Darjeeling
OFFSET 50 — O(n) skip, crawls on deep pages
coldtea.pane_write→ Darjeeling
keyset confirmed — approved
>Run /review on my current changes

Comment like it's Figma

⌘⇧-click any agent output to comment in place. Stack multiple reviews and send at once. Every note reaches your agent anchored to the exact reference

DARJEELING-WT
Claude Codev2.1.160Opus 4.8 (1M context) with max effort · Claude Max~/dev/coldtea
>trim the !important comment — keep only the why
Shortened to 4 lines, focused on the why:
/* !important is load-bearing, not stylistic: the theme boot script writes the palette inline on <html>, and an inline style beats any :root selector — only !important wins. */
Both whys a future reader can't infer survive: why the !important is there (so nobody "cleans it up") and why it's scoped here, not the shared token.
∗ Sautéed for 30s
✻ recap: made the landing background a deeper near-black so the dark pops.
How is Claude doing this session? (optional)
1Bad2Fine3Good0Dismiss
>

Run dozens of agents at once

One click spins up an isolated worktree. Run any coding agent in parallel, each tab tinted so you always know which world you're in

COLDTEA
COL-198×
COL-204×
COL-211×
Worktree
Run
Claude Codev2.1.160Opus 4.8 (1M context) with max effort · Claude Max~/Documents/dev/.coldtea/worktrees/coldtea/darjeeling-worktree-1
>per-worktree accent tint on the active tab
Write(app/worktrees/accent.ts)
+42
Edit(WorktreeModeBanner.tsx)
+8−3
Steeping… (8s · esc to interrupt)
>add a reduced-motion fallback
▶▶bypass permissions on(shift+tab to cycle)

Build at the speed of AI
without shipping bugs to users

In this terminal, your coding agents build, our QA agents verify every PR like a real user then route failures back automatically

Feels familiar

It's still your terminal: same shells, same muscle memory. Just better suited for agentic workflows

Privacy & security

Your repo never leaves your machine. Run background agents in the cloud only when you choose to

No lock-in

Bring your own agents: Claude Code, Codex, Gemini CLI, OpenCode, etc. Coldtea is agent agnostic

Try Coldtea now