Zexet Harness · native Rust runtime

The agent harness you can watch think.

Multi-agent strategies over any model, behind guarded tools and real sandboxes — every plan, tool call, sub-agent and checkpoint visible in one timeline. The cloud runs the native Rust engine, the same zexet-engine binary as your desktop.

  • runtime: rust
  • parallel agents
  • workspace-write
  • bring your own keys

Free locally · no credit card · your keys, your models

Zexet Studio running on the Rust engine: a swarm of three agents working in parallel in workspace-write mode, a capped output stored in the spill store, the context meter and a checkpoint, all in one timeline.

Works with your providers

  • anthropic
  • openai
  • googlegemini
  • mistralai
  • ollama
  • huggingface
  • nvidia
  • perplexity
  • meta
  • 1native engine
  • 9model providers
  • 4multi-agent strategies
  • 3sandbox modes

01 / Runtime

One Rust engine. Locally, and in the cloud.

The guarded agent loop, repetition guard, token budget, cancellation, spill store and live context accounting are native Rust. Cloud workspaces run that same engine — not a port, not a wrapper.

  • Parallel, multi-agent strategies

    Run a goal as single, swarm, debate or hierarchical. A planner fans work out to specialists in parallel; every hand-off lands in the timeline.

    --strategy single | swarm | debate | hierarchical

  • Sandbox modes

    read-only, workspace-write or danger-full-access. An unknown value aborts the boot instead of relaxing; only a human widens it.

    --sandbox workspace-write

  • Spill store

    Long tool output is capped, never lost: it is stored once and handed back as a spill:// id the agent pages through with spill_read.

    spill_read · spill_list

  • context_status

    Live accounting of where the tokens go — system, tool schemas, conversation, tool output — against the model's window and the compaction line.

    used / usable · compaction line

  • Checkpoints & snapshots

    Every run captures a checkpoint and every write a snapshot. Resume a plan or roll the workspace back to any point.

    checkpoint list | restore <id>

  • Guarded tools

    A constitution blocks destructive shell patterns, writes outside the workspace and secret reads, in every mode and every surface.

    constitution · repetition guard

zexet-engine
The harness: agent loop, tools, strategies, Studio server. Serves the Studio in every cloud workspace.
zexet-core
Spill store, sessions and storage, byte-compatible board and notes.
zexet-sandbox
read-only · workspace-write · danger-full-access, enforced per tool call.
zexet-desktop
Native desktop shell around the same engine and the same Studio.

02 / How it works

From goal to reviewed work in three steps.

No DSLs, no scaffolding. Point Zexet at a goal, pick a strategy, and inspect what happened — every step is visible and reversible.

  1. Connect your models

    Install the CLI, sign this machine in with zexet login, and add keys for any supported provider — Anthropic, OpenAI, Gemini, Mistral, Ollama and more.

  2. Run a goal

    Choose single, swarm, debate or hierarchical. The planner fans work out across agents, and every tool call lands in a live timeline.

  3. Review and resume

    Watch checkpoints and snapshots stack up. Approve writes, roll back any point, and let the engine keep working while your laptop is closed.

03 / Run it

Three ways to run Zexet.

Local

CLI · TUI · desktop

Everything on your machine. Build the Rust runtime, sign this machine in to your account, and run a goal.

$ npm run build:rs
$ zexet login
$ zexet run "fix the login bug" \
    --strategy swarm --allow-write

Cloud

rust engine

Your own zexet-engine container in workspace-write mode, with its own network and limits. It keeps working when your laptop is closed.

# from your dashboard
Workspaces → New workspace → Open
# GET /health → {"runtime":"rust"}
Create a workspace

Self-hosted

docker compose

Run the engine on your own server behind TLS. The token opens the page once; then an HttpOnly cookie carries the session.

$ export ZEXET_UI_TOKEN=$(openssl rand -hex 24)
$ export ZEXET_DOMAIN=agent.example.com
$ docker compose -f deploy/cloud/docker-compose.yml \
    up -d --build

04 / Pricing

Free on your machine. Pay for the cloud.

  • Free

    Run Zexet on your own machine. Everything local, no card.

    $0forever

    • CLI, TUI, Studio and desktop app
    • All strategies: single, swarm, debate, hierarchical
    • Bring your own provider keys
    • 1 cloud workspace (trial, sleeps when idle)
    • 2 API keys
    Get started
  • Pro

    Popular

    A managed workspace that keeps working while your laptop is closed.

    $12per month

    • Everything in Free
    • 3 cloud workspaces
    • Checkpoints and snapshots kept in the cloud
    • Job-finished notifications
    • 20 API keys
  • Team

    Shared workspaces and keys for a small engineering team.

    $39per seat / month

    • Everything in Pro
    • 10 cloud workspaces
    • Shared connectors and skills
    • Priority support
    • 100 API keys
Paid plans open soon. Subscribing now puts you on the list and we'll email you when checkout is live — no card needed today.

05 / FAQ

Questions, answered.

The short version of what Zexet is, where your data goes, and how the paid plans work.

What is Zexet?

A composable agent harness on a native Rust runtime. It runs multi-agent strategies over any model, behind guarded tools and real sandboxes, with every step visible in one timeline.

Which models can I use?

Anthropic, OpenAI, Google Gemini, Mistral, Ollama, Hugging Face, NVIDIA, Perplexity and Meta today. You bring your own keys — Zexet never locks you into one vendor.

Where does my data go?

Locally, nothing leaves your machine. In cloud workspaces, your engine runs in an isolated container with its own network, resource limits and a read-only root filesystem; secrets are encrypted on the server.

Is the cloud engine really the same as desktop?

Yes. Cloud workspaces run the same zexet-engine binary you run on your desktop — not a port and not a wrapper — so behavior is identical.

How does the paid waitlist work?

Subscribing while checkout is closed just puts you on the list. No card is required, and we email you when Stripe checkout goes live.

Can I self-host?

Yes — Docker Compose behind TLS, with Cloudflare DNS/proxy and Caddy terminating TLS. The deploy script provisions the VPS and the whole stack.

Start

Start with one goal.

Create an account, connect your machine with zexet login, and watch the timeline fill in.

Create your account