Open-source AI work infrastructure

Let AI actually work on the machines you own.

Stop copying commands, logs, and files back and forth. OpenSynapse gives an AI client bounded access to an authorized machine, lets it do real work, and checks the result.

Public Alpha evidence: fresh public install passed, 59/59 regressions passed, Linux + Android bounded work passed, and a real ChatGPT → OpenAI Secure MCP Tunnel → OpenSynapse write → readback E2E passed.

The human relay is the problem.

AI can already explain what to do. The frustrating part is still you moving commands, error logs, files, and machine state between the chat and the place where work actually happens.

No command copy-paste

Ask for the outcome. The authorized node performs bounded work instead of handing the terminal back to you.

Real result verification

“Done” is not enough. Meaningful actions should finish with an observable check when the target can be verified.

Your machines, one model

Server first. Desktop and Android nodes follow the same public interaction model while remaining independent runtimes.

Ask naturally. Let the node do the work.

1
Install on a machine you ownChoose Linux or Android/Termux in the Getting Started guide, then verify the explicit access boundary.
2
Choose what AI may accessRead roots, write roots, and allowed commands are explicit. Arbitrary shell is not enabled by default.
3
Connect an AI client through MCPThe local node exposes bounded tools through MCP. Real-account ChatGPT → OpenAI Secure MCP Tunnel → OpenSynapse write → readback is verified now.
4
Verify the requested outcomeOpenSynapse is designed around observable results rather than worker self-report alone.

Linux Alpha is public now.

This is an early usable base, not a claim that the whole vision is finished.

git clone https://github.com/nslabhwan/ns-agent-reliability.git
cd ns-agent-reliability
OPENSYNAPSE_ROOT=/srv/my-app bash install.sh
opensynapse doctor

# Optional OpenAI Secure MCP Tunnel prepare path
opensynapse connect openai \
  --tunnel-id tunnel_REPLACE_ME \
  --prepare-only

Verified now

Linux one-install bootstrap
Android/Termux fresh install on real device
Bounded MCP read / write / commands
Android stdlib MCP write + readback E2E
Real ChatGPT Secure MCP Tunnel write + readback E2E
59/59 regressions

What should this feel like?

The target is not another dashboard. It is the AI interface you already use gaining a safe path to real work.

“Check this server and fix the failing test.”

Inspect allowed files, run bounded commands, make the permitted change, then verify the test result.

“Do this on my phone.”

Android itself becomes an independent work node. Android/Termux is now a verified public node. Media and device-control modules are the next Android expansion.

“Continue the job and tell me when it is actually done.”

Long-running resume and continuity are planned around the same verified-result principle.

Power without an open shell to the Internet.

Machine access is a trust problem, so safety has to be visible to the user rather than hidden in architecture diagrams.

Explicit roots Bounded actions shell=false Secret-path denial Atomic writes Open source Self-hostable core\nOfficial tunnel-client SHA256 verified

The current local HTTP endpoint binds to loopback by default. Do not expose an unauthenticated local endpoint directly to the public Internet.

We need real users more than more hidden features.

Try the Linux Alpha on a disposable or non-critical workspace. Tell us where installation, trust, permissions, or the first real task feels confusing. Do not include secrets in public issues.