Coldtea is a desktop app that lets you run the coding agents you already have on hand ─ Claude Code, Codex, Gemini CLI, and OpenCode ─ inside a single terminal environment, either individually or as a team. What sets it apart is that it “doesn’t stop at the pull request.” After an agent writes code, a QA agent clicks through the actual app like a human would and runs regression tests, while a production monitoring agent reads errors and feedback and files broken spots as tasks. The design aims to let agents drive the entire development lifecycle. Launched in August 2026 and based in Germany, the service is primarily aimed at solo developers and small teams who want to run multiple agents in parallel.
Key Features
- Launch four kinds of agent CLIs as they are:
claude/codex/gemini/opencodestart directly in local panes. Coldtea does not replace any of the CLIs; authentication, model selection, slash commands, and permission models all stay with each CLI - Agent teams and parallel work: Assign one pane per role and launch multiple agents at once. If you split working directories with git worktrees, agents can proceed in parallel without fighting over the same checkout
- Plans and logs stay attached to tasks: Import boards from Linear, Jira, or GitHub Issues with one click, with two-way sync. Each session’s plan and execution log is tied to its task, so you can trace what happened even if you come back a week later or hand the work off to someone else
- The QA agent “Matcha” runs regression tests on every PR: Write the user flows you want tested in plain language, then hand over a URL or a build, and Matcha operates the actual app to verify them. Results come as Pass / Fail / Warning verdicts backed by evidence ─ steps, screenshots, logs, and recordings ─ which cuts down on the time people spend re-tracing why something failed
- Mobile app QA is supported too: Verify Android (Pixel 7 / Android 15) and iOS (iPhone 15 Pro Max / iOS 17) builds on Appetize’s cloud devices. Builds can be supplied via file upload or from GitHub Actions artifacts
- A production monitoring agent files bugs for you: Connect Sentry, PostHog, logs, feedback inboxes, and similar sources, and the monitoring agent reads them all day like an on-call engineer, compiling investigated findings into a daily brief and tasks
- Verification happens inside the app: An embedded editor based on OpenVSCode and an in-app browser for your local dev server let you open and check your changes on the spot. Launched panes receive session-scoped MCP access, so agents can read task context and drive the in-app browser themselves
Pricing
The structure makes the parts you use every day free and limits usage-based billing to what you actually consume.
| Feature | Price | Details |
|---|---|---|
| Terminal | Free (Free forever) | No limits on agents, teams, or seats, and no trial expiration |
| Production monitoring | Free (bring your own keys) | Reads monitoring data you already pay for, such as Sentry, PostHog, and Datadog |
| Agent QA | Free up to 2,000 credits per month | Shared across web and mobile tests. Buy additional credits when you run short |
The official site explicitly states “no sales calls, no minimum seats, no contact-us-only plans.” Usage fees for the agent CLIs themselves (a Claude subscription, each vendor’s API keys, and so on) are expected to be arranged by you.
Pricing information is current as of August 2026.
/pricingcurrently returns a 404, so please check the terms on the official site homepage.
Pros and Cons
✅ Pros
- You can keep using the agent CLIs you already know. You are not locked into a proprietary chat UI
- Coding, regression testing, and production monitoring live in the same environment, so “what comes after the PR” isn’t scattered across separate tools
- Plans and logs stay on the task, so you can read back through work you delegated to agents
- QA verdicts come with evidence, making it harder to confuse “a bug in the app” with “a problem in the execution environment”
- Parallel execution of multiple agents and worktree-based workflows are assumed from the start
⚠️ Cons
- Only a macOS build (Apple silicon / Intel) is distributed; Windows and Linux versions are not officially announced
- The pricing structure is not published, which makes cost estimates hard to build
- Coldtea itself adds no sandbox. Local panes run with your own permissions, and worktrees only separate Git working directories ─ credentials and network access are not isolated
- Installing and authenticating agents is your responsibility. If a CLI doesn’t work in a regular terminal, it won’t work in Coldtea either
- For mobile QA, iOS targets Simulator builds only; device-targeted
.ipafiles are not supported - It is a new service launched in August 2026, and its features and stability are still maturing
Comparison With Similar Services
| Criteria | Coldtea | Standalone agent CLI (Claude Code, etc.) | AI-integrated editor (Cursor, etc.) |
|---|---|---|---|
| Form | Desktop development environment | CLI running in a terminal | Editor |
| Agents | Launches and combines 4 kinds of CLIs | Single | Built-in agent |
| Parallel execution | Multiple at once via panes / worktrees | Split sessions yourself | Implementation-dependent |
| QA testing | Built-in QA agent | Separate tool | Separate tool |
| Production monitoring | Built-in monitoring agent | Separate tool | Separate tool |
Who Is It For
- Developers who use Claude Code or Codex daily and want to move toward running multiple agents in parallel
- Teams that can’t keep up with verifying agent-written code and want to automate regression tests per PR
- Small teams that want to reduce the manual effort of picking up production errors and feedback and turning them into tickets
- Developers who want to hand regression testing for iOS / Android apps to agents, not just web and desktop
Summary
Coldtea is a tool that tries to extend the development environment ─ which tends to stop at “have an agent write the code” ─ all the way through testing and production monitoring. Because its design doesn’t replace existing CLIs, you can adopt it without abandoning the way you already work with your agents, and Matcha’s evidence-backed verdicts together with the monitoring agent’s automatic ticket filing fill the “verification” gap that is weakest in agent-driven development. The allocation ─ terminal and production monitoring free, with only QA on credits ─ also lowers the barrier to getting started. The macOS-only restriction is a real constraint, but it’s worth trying the full loop on a small repository first.