03 / AI workspace · Windows / WSL / OpenCode

SeneschalModels, tools and a workspace of your own.

A personal workspace for OpenCode on Windows and WSL. Keep conversations, models, tools and multi-agent work together, with explicit approvals and a configurable interface.

For example, two agents can finish separate tasks before a reviewer receives both results. The examples below show what happens when a prerequisite or access decision holds up the next task.

Understand the workflow

How Seneschal connects tasks.

01 / How tasks connect

Full-size map ↗

Select a box to read its role and connections. Its example link names the step you can watch; some components share an animation.

02 / Watch a task continue

Full-size guide ↗

The carrier is a task result. The barrier represents its access policy. Try a wait or a correction.Repeats automatically while in view. Pause to hold a moment.

These examples explain the workflow without running agents or reading sessions. A script-free map is also available.

Seneschal Night Studio workspace with city pixel art and its desktop mascot
Seneschal Night Studio with city pixel art and the desktop mascot. Screenshot from the Beta 20 repository.

The current workspace / Beta 20

Beyond the task board.

Models, roles and approvals

Choose a model per session. Plan explores without changing files, Build works under the configured permission policy, and Chat leaves active tools off. Tool availability and permission to use it are separate decisions.

Teams with retained context

Turn one selected plan into a team, or configure the board yourself. Up to eight independent workers can run in parallel. Dependencies pass completed results downstream; bounded correction requests and a configurable supervisor keep review visible. Saved boards retain their linked sessions.

Activity you can inspect

Overview, Sessions, Usage and Service views bring activity into the workspace. An optional local history service indexes OpenCode history. Live progress and recorded history remain separate, and usage estimates are not provider invoices.

Voice and a 3D companion

Use browser speech or optional NVIDIA cloud speech, with separate mic and speaker controls. The local companion follows listening, thinking and speaking, with docking, resizing and camera controls. NVIDIA speech requires explicit consent and sends audio to its provider.

A workspace that feels personal

Nine animated wallpaper scenes include the Afterlight worlds, with scene interactions, lighting and motion controls. English and فارسی controls support right-to-left panels without rewriting technical paths or chat content. Wallpaper art needs no model or speech call.

Your paired phone, your running PC

Private Phone Access pairs your own browser through Tailscale after the installation passes its local checks. The PC stays awake and runs the engine. Pairing, credentials and sensitive service controls stay on the PC; this is remote access, not cloud hosting.

Optional integrations connect browser automation, Blender and a VS Code companion. Each needs its own supported local setup. Explore the controls in the user guide ↗

My contribution

Keeping dependent
work connected.

Working with several agents creates decisions about who does what, which tasks depend on others and what an agent may change. I structured the workflow and directed the AI-assisted implementation: sessions, roles, access controls and a board for coordinating dependent work.

A dependent task starts after its selected prerequisites finish. Each board card retains a session, and a configured supervisor can review the completed work. The source and beta releases document the Windows/WSL setup and local integrations.

Workflow decisions

What a task needs before it continues.

Wait for selected prerequisites

Selected dependencies define what a task waits for. A downstream task starts after all its selected prerequisite tasks finish. The overview omits parallel branches and repeated tasks to keep the main path readable.

Choose the access policy

Modes and permissions describe the selected agent’s access. Plan mode is read-only; Build asks by default. The policy is configurable, and actions, providers and integrations do not all have the same enforcement.

Keep the result with its session

Each board card is an OpenCode session. Completed prerequisites can pass labelled handoffs downstream, keeping the context and result connected to the dependent work.

Review before acceptance

A configured supervisor is responsible for final acceptance. The board makes review and handoffs visible. Generated work still needs to be checked for correctness and safety.

Download and setup

Start with the published beta.

Windows / WSL setup

Requires Windows 11, WSL, Node.js 20+ and an existing OpenCode installation with a connected provider. Node 22.16+ is recommended for the full feature set. Download the Beta 20 ZIP, extract it and run install.cmd; then open the desktop shortcut. It is a script-based installer, not a standalone signed native app.

Updates and private access

Packaged copies check published releases through Settings → App updates. Downloads are checksum-verified and local data is preserved. Source checkouts update through Git. Finish active tasks before restarting. Phone Access has its own setup and verification steps.

What you can inspect.

Published Beta 20

The release, user guide and current repository document the workspace, Agent Board, Activity, voice and phone access. The screenshot below the interactive examples matches the image in the current repository.

Published validation scope

Beta 20 reports 288 passing source/package tests and 207 separate synthetic Phone Access checks. Those checks cover source behavior and packaging; they do not replace testing a fresh Windows/WSL installation, physical phone or microphone. Read the validation record ↗

Beta and privacy boundaries.

Cloud model use depends on the selected provider. A local workspace does not mean that model requests remain entirely on the computer; the provider receives the prompt context needed for a request.

This page runs no agents and reads no private sessions. Its interactive map and railway explain the task-board workflow, not every feature listed above. Beta software and generated results still require review.

Updated 1 November 2026 against source revision 9758760 and the published Beta 20 release.