Jon Oringer started Shutterstock in 2003 and led it as CEO through its 2012 NYSE listing.
Now he builds Sandboxer at North Bay Ventures, writing product and working directly with the AI agents that do much of the engineering and operations work.
https://t.co/bSfxv1d9vU
Run n8n in a sandbox with webhooks that work: its Docker image with a volume, N8N_HOST, N8N_PROTOCOL and N8N_WEBHOOK_URL set to the hosted address, then sbx host 5678. The URL is private by default.
https://t.co/Kydzmgwvq5
What teams use Sandboxer for:
- coding agents
- many agents in parallel
- untrusted or AI-generated code
- evals and tests in clean environments
Agents drive it through SDKs, a CLI, a REST API or a hosted MCP server.
https://t.co/2q26Zlcwjq
Want every coding agent to behave alike? Harnesses read their usual config from home: CLAUDE.md, AGENTS.md, MCP configs, skills, settings.
Commit them to your repo, or bake them into a snapshot:
sbx snapshot <id> agent-ready
sbx new --from agent-ready
https://t.co/VLNCTnRm7b
Did the agent's change actually help? Fork one prepared sandbox and score the unchanged code and the change on the same disk, several rounds each, interleaved. Count a speed change only if it beats the baseline's own spread.
https://t.co/kKyGTUGcf7
What does an unknown GitHub repo really do? Run it in a fresh no-env sandbox: pin one commit, run the README's install and quickstart, record each step's exit code, copy the report out, delete the VM. Your laptop and keys stay out of reach.
https://t.co/YFBuOtfoWC
Every way of starting a service survives a resume that restores memory. After a cold boot, a reboot or a killed process, only an enabled systemd unit with Restart=always, or a container with --restart unless-stopped, comes back.
https://t.co/XviSafzSRS
Make a leaked key do little harm. Scope it to one sandbox, a few actions and a week:
sbx api-key create one-sandbox --sandbox <id> --actions https://t.co/0Ox2ftlXNC,exec,https://t.co/lXWzv6rvXR --ttl 7d
https://t.co/GotaMY4mcY
From CLI to code in two steps:
sbx api-key create my-app --preset ci
export SANDBOXER_API_KEY=<the key it prints>
Then POST /api/v1/sandboxes. It returns 202 while the machine boots; the sandboxer-sdk packages for TypeScript and Python wait until ready.
https://t.co/mXeaBeQ3fb
SEO wins a spot in a list of links. AEO wins the passage an AI assistant quotes: one question per page, the answer first, exact facts dated when checked, and crawlers let in.
Sandboxer's guides open with a short answer and date their measurements.
https://t.co/mxtRVpb99V
Ollama on CPU in a sandbox: install with its Linux script, pull a small model, bind to 0.0.0.0, then sbx host the port for a private HTTPS URL. There's no GPU; on Oct 27 we saw about 30 to 55 tokens per second for a 0.5B model and 13 to 22 for 1B.
https://t.co/ZQ76yUg8VX
An open port doesn't mean a ready service. In our Oct 27 test, Grafana's port opened at 0.4 s but answered at 3.1 s. Poll a real health URL with a timeout instead, and from outside, treat a 502 from the hosted URL as "not listening yet".
https://t.co/Pq4DNMPBvz
A microVM narrows the escape routes. It doesn't make everything safe. Still your job:
- secrets you put inside: code in the VM can read and send them
- outbound network: a misled agent can call out
- the output: review code before you merge or deploy it
https://t.co/TtAaYtmNkR
What should never go inside an agent sandbox? Anything you wouldn't hand to the agent's worst input:
- your SSH keys
- long-lived cloud credentials
- production database access
- personal browser sessions
Pass only short-lived, narrowly scoped keys.
https://t.co/HtVNLRO0vt
https://t.co/dR1yd1FY8K isn't Flux Flagger. Flux Flagger shifts Kubernetes traffic between versions. https://t.co/dR1yd1FY8K is a hosted feature-flag service: agents create, roll out and kill flags over MCP, a CLI and an API, and every write needs a note.
https://t.co/wWH10tnafm
Writing an MCP server? The model picks tools by name and description alone. Name each after its action, say in one line what it does and returns, keep inputs few, offer only the tools a connection may use, and make writes leave a trail.
https://t.co/ae997NISIR
Want a human yes before an agent runs a risky command? Give the agent no Sandboxer key, only a tool that calls a gate you run. Allowlisted read-only commands go straight through; anything else waits for approval.
https://t.co/iA8T4mQ1f1
Opening a private sandbox URL with its token sets a 30-day cookie, so assets and hot reload just work in the browser. Treat the link and cookie like a password.
Switching a port between --public and --private mints a new token, so old links stop working.
https://t.co/CCrfwhD6Pc