Apache-2.0 agent runtime

The coding agent you can govern.

Rules your repository enforces, budgets that are walls, and a byte-exact record of every run. Read the loop, change it, replay it offline, and test its behavior like code.

Download Jern Read the docs

A terminal demonstration of Jern policy, deterministic golden checks, a caught behavior change, and a run receipt.

Real output from the public demo repository. The policy and replay checks run offline without calling a model.

Use · govern · edit · test

Use it

Anthropic, OpenAI, Ollama, or any OpenAI-compatible endpoint, with streaming and git-safe edits.

Govern it

Scope edits, allow commands, deny tools, and compose protected restrictions that later layers cannot weaken.

Edit it

The host that talks to providers, confines tools, and writes the evidence is F#. The agent that decides what to do next is IronKernel source you can read and edit, and change the control flow without recompiling the host.

Test it

Replay recorded model and tool effects byte-exactly and assert properties of the entire trajectory.

Install

$ curl -fsSL https://jern.ai/install.sh | sh
$ export ANTHROPIC_API_KEY=…
$ jern run "fix the failing test"
$ jern policy
$ jern receipt --md

Why the runtime matters

Agent code runs in a capability environment with no ambient file, process, network, or model authority. Every effect crosses the handler stack where policy, approvals, budgets, and tracing apply. Process-level tools still require an honest OS-sandbox and approval story; the security model documents the boundary and its limits.