Claude Can Now Find Its Own Dependencies and API Keys—Key Custody Is the Real Problem

Claude Code's cloud sessions in Pro, Max, and Ultra plans have one more step: it first figures out the project's dependencies and the API keys needed to run, then starts the session. The main video posted by TestingCatalog demonstrates this process.

Previously, getting from repo to running was blocked by environment setup, missing dependencies, and finding keys. One reply said the setup step usually wastes the most time. Now Claude takes over this part, saving a few detours. It’s available across Pro, Max, and Ultra, a wider rollout than expected.

On cost, a reply mentioned that cloud VMs are not charged separately and count toward plan usage limits. For heavy users, that is more substantial than adding a button.

But the feature description did not answer key questions. Chahat Sharma asked directly: the clip does not say whether the key is read from a local .env or must first be configured in cloud secrets. Randalone asked whether it will ask before using the API key it found. Elias wanted to see exactly which keys it chose and what those keys can access. A cloud session holding production keys is as dangerous as a laptop holding production keys—just easier to overlook.

A reply mentioned Anthropic's statement: new cloud sessions will store API keys for reuse but will not let keys enter the model context. This isolation claim needs verification. Around the same time, someone dug up the issue of Claude Code fixing account email entering request headers, saying v2.1.317 used a 171-line regex, and independent testing still had 8 leak forms pass after 23 attempts. That number cannot be verified from the source material, but it shows the trust point is not in the demo.

Another reply was more practical: after a cloud session finishes, grep the generated scripts for /home/claude paths. Their sandbox path leaked into build files, causing CI to fail on other machines.

There were constructive suggestions too. INNERLUXES said the interaction pattern for network allowlists is right, and keys should be handled the same way. Most .env files give sessions far more permissions than the task needs; issuing short-lived, least-privilege keys per task is better than stuffing the whole .env in. Tink's judgment was more direct: automatic dependency installation is only convenience; the threat model changes when keys do not enter the agent filesystem.

One reply joked that it would look for keys in 2022 Slack DMs like every developer.

ShadowAguy's reply included an image, saying that cloud sessions on paid plans can handle dependencies and keys on their own, making agent evaluation feasible.

Someone in the replies also shared a 28-minute video. TheEasySale placed Claude Code's value in subagent parallelism and task decomposition, with writing code just one part of it. The more finely small tasks are broken down, the less rework. This view comes from a reply, not an official statement.

The feature itself is not complex; what is complex is the permission boundary. Now that Claude can find keys on its own, the next step should be letting people see at a glance which key it picked up and which door it can open.

Original tweet: https://x.com/testingcatalog/status/2118844509346005399