Run untrusted code inside a boundary you control.
jhansi is a self-hosted runtime for untrusted and AI-generated code — simple to run, provable by design. It runs on infrastructure you control, with a durable record of what was authorised and how the execution ended.
Generating the code is the easy part.
Generated code touches files, networks, credentials and host resources.
Running it inside your perimeter doesn't make the execution bounded or provable.
They aren't a trustworthy account of what was authorised and what actually happened.
Between the agent and your infrastructure.
AI agents can generate the code. jhansi gives that code a controlled place to run — and records what happened while the execution is taking place, rather than reconstructing it afterwards.
jhansi · execution boundary
Simple to run. Provable by design.
For developers building agent workflows, platform teams defining an execution standard, and regulated teams that need the boundary to stay inside their environment.
Self-hosted, complete on its own.
One machine is a first-class deployment, not a cut-down cluster install. The direction is a single binary you install and start.
No Kubernetes, no Helm and no mandatory cloud control plane for the core product.
Evidence designed into the engine.
Structured events come off the execution lifecycle itself — including the transitions that fail and the ones that get denied — rather than being reconstructed later.
That provides a durable execution history now and a path toward exportable, tamper-evident evidence as the runtime evolves. These are controls and evidence that can support compliance work — not a compliance claim.
A record, not just a log file.
Every state change produces a structured event, including failures and denied transitions. The record is created as the execution moves through its lifecycle.
The current design provides an append-only execution history. Export and tamper-evidence capabilities will evolve with the runtime.
Shape it before it ships.
jhansi is being rebuilt in Go around the core create → execute → destroy lifecycle, with the evidence spine designed into the foundation. Nothing to install yet — the execution model, evidence format and isolation boundaries are being developed in public. Join the discussion on GitHub or follow progress on the roadmap.