You dream it. We build it together.
Halyard is the frontier agentic coding harness: a local-first runtime on a Rust kernel. Run one agent or a whole fleet unattended for minutes or hours, with a receipt for every step. You set the rules and the risk you'll manage. You dream it up, and we build it together.
What a harness with its own kernel gives you.
Runs at the risk you set
Work below your line runs on its own, for minutes or hours. Steps above it come to your phone, wherever you are. Answer there and the work keeps going.
Approved from iPhone · agent resumed
A receipt for every step
The kernel logs every tool call and checkpoints every action, so you can review, revert, fork or replay the work. Steps that reach outside the workspace are marked as ones Halyard can't undo.
Restored app.ts to how it was before this action
Sized before it runs
Big jobs are surveyed first, and the agent count and estimated cost show at a gate before anything runs.
12 agents across 3 machines · estimated $4.10 · waiting for you
How Halyard differs from the tools you already run.
Claude Code and the Codex app are good agents. Conductor and Herdr run several of them side by side. Halyard is the layer underneath: its own kernel decides what runs without you, sandboxes it, and keeps the receipt.
| Halyard | Claude Code | Codex app | Conductor | Herdr | |
|---|---|---|---|---|---|
| What it is | A runtime with its own Rust kernel | Anthropic's coding agent | OpenAI's coding agent | A Mac app that runs other agents side by side | An open-source runtime that hosts other agents' CLIs |
| Runs on | macOS 14+, Windows, Linux | macOS, Linux, WSL, Windows | macOS, Windows, Linux | Mac | macOS (Apple silicon), Linux, Windows |
| Several agents at once | One agent or a fleet, across this Mac*, Anthropic and OpenAI | Subagents, background sessions, cloud sessions | Subagents, a worktree per chat, cloud tasks | A workspace and branch per task | Agents in terminal panes, on any machine over SSH |
| What runs without asking | A risk level you manage; the kernel enforces it at each step | Permission modes with allow, ask and deny rules | Permission modes, including “Approve for me” | — | Left to each agent's CLI |
| Sandbox | Every agent, always | OS sandbox for shell commands, off by default | OS sandbox with read-only and workspace-write modes | — | Through wrappers or a Vercel Sandbox plugin |
| Undo | A checkpoint for every action, not just every turn; actions it can't undo are marked | Checkpoints per prompt; shell changes not covered | — | — | — |
| Receipts | An audit log of every tool call, on your machines | OpenTelemetry events, opt-in | Audit events for enterprise plans | — | — |
| Models | Any model with your own keys, or via local models | Claude, via Anthropic or your cloud provider | OpenAI models | Whichever agent you run in it | Whichever of 20+ agent CLIs you run |
| Sized before it runs | Agent count and estimated cost at a gate | — | — | — | — |
| From your phone | Halyard app on iPhone and Android | Remote Control and the Claude app | ChatGPT mobile app | — | iPhone, iPad and Android apps |
— means the vendor's own pages don't mention it, not that the feature is missing. Checked against their docs in November 2026.
* Add additional machines as runners.
The agent works. When a step is risky, it stops and asks you.
This run was generated for you from the page's seed. The plan streams in, then the kernel holds the risky step until you decide. Approve it, deny it, or stop the agent. Afterwards you can revert the whole turn.
- Read the orders handler and middleware
- Add migration 0042_idempotency_keys
- Implement the middleware and Postgres store
- Run the API tests
- Apply the migration to staging Needs you
- Hand the change over for review
One kernel on your machine. The apps are windows onto it.
The kernel runs agents in a sandbox against your repositories, keeps a checkpoint for every action and writes the audit trail. The desktop and phone apps talk to it through its protocol, so either one can approve, deny or stop.
Native apps on desktop and phone, talking to the same kernel.
Plan, watch and decide at your desk. When you step away, the app shows the same approvals and the same stop button.
Put Halyard on your machine before it ships.
We are letting in a small group of engineers who already run coding agents and want to see how far they can go.
- macOS 14 or later on Apple Silicon or Intel, Windows and Linux
- Bring your own model, or your own keys
- Runs against your repositories, on your machine, with supervision from your phone
- A direct line to the people building the kernel