Work

Systems we’ve built, and the pattern behind them.

We build for capability and safety at the same time: agents that do real work, and can’t do the wrong irreversible thing without a person saying yes.

Flagship · on-device · human-in-the-loop

LocalAgent, a private multi-agent runtime

A complete agent system that runs on one machine: a local model plans and drives real tools, hands the hard parts to a stronger model, and gates every risky action behind a human. Built end to end as our reference architecture.

reason

Local model

Private and on-device. Plans, calls tools, self-corrects with anti-loop guards.

act

Tool layer (MCP)

Files, shell, search, web. Reads run free; state changes are gated.

approve

Human gate

Every risky action waits for a one-tap approve or deny. Full audit trail on disk.

escalate

Model hand-off

Stuck work routes to a stronger agent that finishes it, on your approval.

Python orchestratorMCP tool servermodel routing approval gatingaudit loggingnative app SSE streaminglaunchd services
The LocalAgent runtime pausing for human approval before issuing a refund

A run in progress: reads happen automatically, but issuing a refund waits for a human. Try it live →

The problem

Off-the-shelf agents are impressive in a demo and dangerous in production. They send your data to a third-party API, they spin in circles on multi-step tasks, and when they do act, they can take an irreversible step, delete a file, move money, email a customer, with no one watching. For a lot of teams that rules agents out entirely.

The build

LocalAgent keeps the model on the machine, so prompts and data never leave. A planner-executor loop drives a typed tool layer over the Model Context Protocol, with anti-loop detection so it escalates instead of grinding. Read-only tools run freely; anything that changes state, or reaches a stronger model, stops for an explicit human approve or deny. Every decision is written to an audit log.

YOUR INFRASTRUCTURE · DATA NEVER LEAVES Local model reason Tool layer act · MCP Human gate approve / deny Stronger model off-machine on approval text only
The model, tools, and gate run inside your walls. Only approved, text-only work ever reaches an outside model.

Capable enough to do the work. Safe enough to let it.

What it demonstrates

This is the pattern we bring to client work: real tool use, private by default, and a human gate that’s a hard control rather than a hopeful line in a prompt. The same architecture adapts to your stack, your tools, and your definition of “risky.”

Open source

Gatekeeper

We open-sourced the core idea: a human-approval gate any MCP agent can call.

A human-approval gate for any MCP agent

Drop-in tool that makes the gate a call the agent can’t route around: before an irreversible action, a person clicks Approve or Deny. Deny-closed by design, if no human can be reached, the answer is Deny, never a silent yes. Native dialog, zero heavy dependencies.

Available for client builds →
🔧

Why it matters

Most “human in the loop” is a sentence in a system prompt the model ignores under pressure. Gatekeeper makes it structural.

See it move

Watch an agent work, with a human gate.

An interactive demo: the agent runs a real workflow and pauses for your approval on every irreversible step. Try denying one.

Open the live demo →