Jern Cloud docs

FAQ and limits.

Short answers, and the numbers behind them.

Limits every attempt runs under

Limit	Value
Tokens per attempt	The session cap, up to the workspace maximum an admin sets under the platform ceiling
Agent phase	The workspace's time limit per attempt, 30 minutes by default, up to 45 minutes on trial, 4 hours on paid plans
Whole machine	The time limit plus setup and publication; a guard of 25 hours behind it
Setup command	10 minutes
Processes per attempt	128
Agent output	10 MiB per phase
Attempts per task	3
Machine	Shared 2 vCPU with 2 GB by default; 4 or 8 dedicated vCPU with 8 or 16 GB by choice, metered per hour
Evidence retention	90 days by default, adjustable

Can the agent merge?

No. It pushes to the session's branch and opens or updates one pull request. Merging is yours.

Can the agent reach the internet?

Under the Open profile, the default, it can reach any public host over HTTPS through the runner's proxy, which refuses private addresses and counts every connection on the receipt. Under a governed baseline it reaches only the Jern gateway and the exact public hosts the baseline opens (network_allow). A setup command may reach the package registries the environment declares, in a separate domain that is destroyed before the agent starts. It can read the repository's issues through the run-scoped MCP server: the control plane fetches them with a read-only token that never enters the machine, and every read is on the receipt.

Does the agent see my API key or GitHub token?

No. The gateway holds your provider key. The checkout token is erased right after the fetch, and publication uses a token the runner obtains only after the policy checks pass.

Where does my code go?

To a machine created for the attempt and destroyed afterwards, and, as prompts, to the model provider under your own key. Evidence is encrypted before it reaches Jern's database. Conversation text is stored in plaintext so the dashboard can show it, and is redacted after the retention window.

Which languages and environments?

Python, Node, .NET, and Java environments are available today, plus a general image with Python, Node, and a native toolchain. The .NET environment carries the .NET 10 SDK and restores from NuGet during setup; the Java environment carries Temurin JDK 21 with Maven and Gradle. See Environments. The environment is pinned per session by revision, and an admin can assign a different one per repository.

Why did my attempt fail?

The attempt card states the reason: cap exhausted, policy refusal, provider rejection, or a publication problem. An attempt that reaches its time limit is not a failure: what the agent had is published and the attempt is marked timed out. The receipt and trace have the detail. Three failed attempts mark the task exhausted.

What does it cost?

The control plane is priced per workspace per month after a 30-day free trial that includes $5 of inference on two open-weight models on our own key, and can be bought through GitHub Marketplace, billed by GitHub to the account that installed the App. Model usage is billed by your provider on your own account. See pricing.

Where do I report a security issue?

security@jern.ai, or GitHub private vulnerability reporting on the open-source repository. See the security page.