Appearance
Vistara
Vistara is the coordination layer for the agents you already run. You keep using Claude Code, Codex and Cursor in the terminals and editors you already use. Vistara runs alongside them on your Mac, keeps track of what each session is doing, and tells you when one of them needs you. You answer from a small card near your cursor, and the answer goes back to that exact session.
Underneath the menu-bar app is a local daemon. It reads your agents' hook events and transcripts, keeps a live state for each session and project, and holds an agent's question or approval request for a short time so you can answer it without switching tabs. Apart from a daily anonymous "installed" ping (which you can turn off), nothing leaves your Mac unless you turn on your node for points, and even then only verification results are sent, never code, prompts or transcripts.
Beta
This is a beta for Apple Silicon Macs. These pages describe what the beta does today. Anything not shipped yet is marked Planned.
Who it is for
- People who run more than one coding agent at a time, in different terminals, tabs or projects.
- People who lose minutes to agents that finished or asked something while they were looking elsewhere.
- Teams that want to hand work from one agent, or one person, to another without re-explaining it. Team features are planned; single-user handoff between agents ships today.
The four primitives
Vistara is built around four ideas from the whitepaper. The beta ships a first, narrow version of each; the table says exactly how narrow.
| Primitive | Question it answers | What ships today | Planned |
|---|---|---|---|
| Workstate | What is happening? | A per-session and per-project state derived from events: status, test and build checks, last change, open question, next step. The ⌘⇧Space overlay shows it. Send work turns it into a handoff another agent can continue from. | The full Workstate record: objective, decisions with authors, constraints, artifacts, failed paths, dependencies, provenance. Sharing between people. |
| Attention | What needs a person? | Cards near the cursor when Claude Code or Codex asks a question or needs approval. Your answer goes back through the agent's own hook. Finished turns go to a quiet list, not a popup. | A brief ranked by risk for when you are away. Attention rules you set in a mandate. |
| Mandate | What may agents do and spend? | A local, unsigned policy file (~/.vistara-relay/policy.json). By default it approves routine reads, edits, tests and builds inside the repo and sends everything else to you. | One signed mandate covering scope, output rules, network, budgets, machine use, decisions and attention, enforced on every action. |
| Verification | Is the work right, checked independently? | Test and build results taken from the agents' own runs. When an agent stops with changes in a git repo, Relay also re-runs the repo's checks in a separate git worktree and looks for weakened tests. | Goals with acceptance checks written before the work starts. Verification levels V0 to V3 in a separate process, including review by a model from a different vendor. Signed receipts. |
Pages
- Getting started: install, the welcome, your first ten minutes, troubleshooting.
- How it works: the daemon, hooks, inbox sockets, the Workstate rules, how replies are routed.
- Send work and handoffs: handing a session's state to another agent.
- Agents and integrations: Claude Code, Codex, Cursor, Warp, Claude Desktop, and the Vistara MCP server.
- Points and Vistara Node: what the node is, how points work, what is sent.
- Privacy and security: what Vistara reads, stores and sends, and the permissions it asks for.
- FAQ
- Roadmap
- Whitepaper summary