OpenCode fork with workspaces, MCP UI, and more. Made for custom agents and harness development. https://docs.vski.ai
  • TypeScript 97.9%
  • CSS 1.8%
  • JavaScript 0.2%
Find a file
Repository files (latest commit first)
Filename Latest commit message Latest commit date
Anton Nesterov 54242ea4e1 feat(extensions): add cross-workspace catalogs and self-improvement tools
Add catalogs that let a user reuse extensions already configured in another
workspace instead of rebuilding them. Managed plugins, MCP servers, classifier
entries, and self-improvement settings can be imported into the current scope;
classifier and MCP copies preserve their credential identity so an existing
sign-in or API key is reused. Imports write only the target workspace.

Add the self-improvement workspace: Settings → Extensions → Self-improvement
now owns collection, opt-outs, schedules, classifier labels and instructions,
judge configuration, sweep logs, and a proposal inbox. Sweeps run inside the
server, catch up after downtime, continue backlogs, and skip until a decision
model is ready. The judge is a chunked stateless call with an allowlisted set of
targets, an editable prompt, and a model selector; approval mode queues
proposals and automatic mode applies them with backups.

Proposal cards open a prefilled review chat in the proposal's own workspace.
The titlebar badge shows pending proposals from every workspace and refreshes
from global events plus a slow fallback poll.

Every sweep now materializes the collected evidence under
improvement/datasets/latest as dataset.json and incidents.jsonl, so the evidence
folder always contains a downloadable dataset. The panel exposes the latest
dataset path.

Add migrations for improvement runs, proposals, and the proposal source
directory. Verified with bun run check (22/22), improvement and classifier
tests, and HTTP smoke tests for settings, catalog/import, and sweeps.
2026-11-17 17:03:16 +01:00
.husky fork opencode 2026-10-28 13:05:11 +01:00
.opencode chore: strip upstream infra, opencode services, telemetry, and the TUI 2026-10-31 00:25:01 +01:00
.zed fork opencode 2026-10-28 13:05:11 +01:00
docs feat(extensions): add cross-workspace catalogs and self-improvement tools 2026-11-17 17:03:16 +01:00
fork feat(extensions): add cross-workspace catalogs and self-improvement tools 2026-11-17 17:03:16 +01:00
packages feat(extensions): add cross-workspace catalogs and self-improvement tools 2026-11-17 17:03:16 +01:00
patches fork opencode 2026-10-28 13:05:11 +01:00
perf fork opencode 2026-10-28 13:05:11 +01:00
script fix(release): upload the update feed last and explain proxy 413s 2026-11-06 23:06:21 +01:00
sdks/vscode fork opencode 2026-10-28 13:05:11 +01:00
specs/v2 fork opencode 2026-10-28 13:05:11 +01:00
.dockerignore fork opencode 2026-10-28 13:05:11 +01:00
.editorconfig fork opencode 2026-10-28 13:05:11 +01:00
.gitattributes fork opencode 2026-10-28 13:05:11 +01:00
.gitignore fork opencode 2026-10-28 13:05:11 +01:00
.gitleaksignore fork opencode 2026-10-28 13:05:11 +01:00
.oxlintrc.json fork opencode 2026-10-28 13:05:11 +01:00
.prettierignore fork opencode 2026-10-28 13:05:11 +01:00
AGENTS.md fix(personalization): give prompt dialogs a scrolling body and fixed footer 2026-11-16 02:53:01 +01:00
bun.lock feat(documents): add a shared markdown document tab 2026-11-14 21:24:52 +01:00
bunfig.toml fork opencode 2026-10-28 13:05:11 +01:00
flake.lock fork opencode 2026-10-28 13:05:11 +01:00
flake.nix fork opencode 2026-10-28 13:05:11 +01:00
LICENSE feat(desktop): linux x64 build + forgejo release scripts 2026-11-06 22:09:35 +01:00
package.json feat(desktop): linux x64 build + forgejo release scripts 2026-11-06 21:51:22 +01:00
README.md feat(voice): read agent responses aloud 2026-11-10 14:59:08 +01:00
sst-env.d.ts fork opencode 2026-10-28 13:05:11 +01:00
tsconfig.json fork opencode 2026-10-28 13:05:11 +01:00
turbo.json chore: drop .github, fork is not hosted on GitHub 2026-10-30 22:53:55 +01:00

VSKI Harness logo

VSKI Harness

An AI coding agent based on opencode, maintained independently.


About this project

VSKI Harness (or just Harness) is an independent fork of opencode, the open source AI coding agent.

We are not associated with the opencode authors. This project is built and maintained independently, is not endorsed by or affiliated with the opencode team or Anomaly, and is not published by them. "opencode" and the opencode marks remain the property of their respective owners; we use the name only to describe derivation, as the upstream license and its trademark guidance require.

We track upstream closely and merge upstream dev regularly, so the core agent stays compatible while we keep our own features and direction. Every deviation from upstream is recorded in fork/features/: one entry per feature, with the instructions used to reconcile upstream merges.

Compatibility is preserved where it matters: configuration files, state paths, environment variables, and the command-line interface keep working the way opencode users expect.

What this fork adds

Workspaces

Sessions are organized as workspace, project, session. A workspace is a named bundle of plugins, MCP servers, skills, instructions, commands, and agents, and comes with two built-ins: default for plain chat with zero tools, and coding for the full agent behavior. Custom workspaces are defined in the spaces config map or created from the in-app dialog.

Tabs, sessions, project membership, and extension settings are all scoped per workspace, and the set of projects belonging to a workspace is durable server state stored in the local database, so it survives client resets and is shared across clients.

Extensions manager

A full-screen extensions view manages extensions in the standard opencode config format:

  • MCP servers: create, edit, and remove, with JSON import and export
  • Skills: create, edit, and remove skill files, with a SKILL.md editor
  • Plugins: read-only for now, until versioned plugin installation lands (see roadmap)

The titlebar's real workspace switcher and project tabs scope the view: a selected project tab edits that project's own opencode.json, no selection edits the active workspace, and the home directory acts as the global/server scope shared by every workspace and is also editable from Settings. Every write goes through the normal config reload, so new sessions and the next prompt of running sessions pick the extensions up without a server restart.

Connections view

A full-screen connection manager replaces the sidebar servers section and the old Settings connection management. Saved servers are listed with live health, the focused server can be switched with one click, and connections can be added, edited, removed, or re-authenticated. The app opens on the login screen whenever no configured server accepts our credentials.

Interactive MCP widgets

MCP servers can ship widgets: UI templates rendered inside the session timeline that can read MCP resources and call MCP tools on the user's behalf.

Plugin manager

Claude-format content plugins (.zcode-plugin/.claude-plugin manifest, skills, commands, agents, .mcp.json) install from git URLs with #ref pins, zip/tar archives, local paths, or a user-maintained local registry in Claude marketplace.json format — no hosted marketplace. The opencode plugin CLI and the extensions manager install, version, update, enable, and remove them; installed components are namespaced (vski:docs, vski/docs) and hot-reload. Private sources authenticate through your git credentials and {env:VAR} tokens that never touch state files. opencode plugin import claude|codex migrates existing Claude Code or Codex setups (plugins, MCP servers, skills, prompts).

Roadmap

Planned next:

  • Workflows plugin
  • Web tools: agent tools for fetching and working with web content
  • Finish the soft rebrand: in-app text says Harness and opencode art is placeholder-blocked; icons, favicons, and CLI help text remain, specified in fork/features/

Recently shipped beyond the feature list above: the terminal UI is gone (the CLI keeps serve, run, and the management commands; the product is the desktop and web app), upstream's deploy and release infrastructure is deleted, and opencode's paid and first-party services, telemetry included, are stripped; the in-app text rebrand to Harness is done and opencode art renders as transparent placeholders until fork branding exists: no Console or Go sign-ins, no Zen provider, no usage limits, no update channel, and no crash reporting.

Everything listed under "What this fork adds" above is shipped.

Desktop hotkeys on Linux

The desktop launcher accepts --show, --hide, and --toggle. When the app is already running, the flag is routed to the running instance and switches window visibility instead of starting a second app; when it is not running, the app simply starts. That makes the flag the thing to bind to a system hotkey — Electron's global shortcuts don't work under Wayland compositors, and this routes through the app's single-instance lock instead. The tray icon's Show / Hide / Quit menu covers the same actions with the mouse.

The command to bind is whatever launches the app:

  • .deb install: /opt/Harness/vskiharness --toggle
  • AppImage: /path/to/harness-desktop-linux-x86_64.AppImage --toggle

--toggle hides the windows when one is on screen and shows them otherwise (a minimized window counts as hidden); --show and --hide force a direction.

Voice input

The launcher accepts voice flags too, so a system hotkey can drive the composer while Harness is in the background:

  • --voice-toggle starts recording; the next press stops, transcribes, and sends. Outside a chat it opens a new chat first.
  • --voice-start starts recording in the current chat.
  • --voice-stop stops and sends the current take.
  • --voice-cancel discards the take or cancels transcription.
  • --voice-new opens a new chat and starts recording there.

The transcription goes to the destination chosen in Settings → Voice & Devices: chat by default, or insert into the prompt. These flags act on the running app; when Harness is not running, the flag only launches it, exactly like --toggle. If another window was last focused, that window receives the command.

Most desktop shortcut systems report only a key press, so --voice-toggle is the single-key push-to-talk control. Where the compositor can bind key release, use --voice-start on press and --voice-stop on release for true hold-to-talk.

See docs/desktop-voice-hotkeys.md for the full guide, including niri, GNOME, Hyprland, sway, and KDE examples, destination behavior, and troubleshooting.

Voice responses

Harness can also read a completed answer aloud through the text-to-speech engine in Settings → Voice & Devices. It is manual by default:

  • /speak reads the last completed response; run it again to stop.
  • /handfree turns automatic reading on or off for the current chat.
  • --speak-toggle reads the last response or stops the reading.
  • --speak-last always starts reading the last response.
  • --speak-stop stops and cancels the reading.
  • --speak-handfree toggles hands-free for the current chat.

Hands-free is per chat, so a long background agent does not start talking unless that chat opted in. While a response is prepared or read, a control in the composer's action row takes the place of the microphone and send button, with pause/resume and stop. The response-ready desktop notification also reads the answer when clicked, and the launcher flags use the same single-instance route as the window and voice input flags without stealing focus. The in-app command can be assigned a shortcut in Settings → Shortcuts.

See docs/voice-responses.md for setup, styles, privacy, Linux binding examples, and troubleshooting.

niri

In ~/.config/niri/config.kdl — spawn takes each argument as its own string and runs without a shell, so use absolute paths:

binds {
    Super+Alt+H { spawn "/opt/Harness/vskiharness" "--toggle"; }
    Super+Alt+V { spawn "/opt/Harness/vskiharness" "--voice-toggle"; }
    Super+Alt+S { spawn "/opt/Harness/vskiharness" "--speak-toggle"; }
}

niri reloads the config as soon as you save it; niri validate checks the file from a terminal.

Ubuntu (GNOME)

From the UI: open Settings → Keyboard → View and Customize Shortcuts…, scroll to Custom Shortcuts, click Add Shortcut, and fill in a name (Toggle Harness), the command (/opt/Harness/vskiharness --toggle), then click Set Shortcut… and press the combination — for example Super+Alt+H.

From the terminal:

path=/org/gnome/settings-daemon/plugins/media-keys/custom-keybindings/custom0/
gsettings set org.gnome.settings-daemon.plugins.media-keys custom-keybindings "['$path']"
gsettings set org.gnome.settings-daemon.plugins.media-keys.custom-keybinding:$path name 'Toggle Harness'
gsettings set org.gnome.settings-daemon.plugins.media-keys.custom-keybinding:$path command '/opt/Harness/vskiharness --toggle'
gsettings set org.gnome.settings-daemon.plugins.media-keys.custom-keybinding:$path binding '<Super><Alt>h'

If you already have custom shortcuts, check gsettings get org.gnome.settings-daemon.plugins.media-keys custom-keybindings and add the new path to that list with the next free customN — the first command above replaces the whole list. Add a second shortcut with --voice-toggle in place of --toggle for voice input, and a third with --speak-toggle for voice responses.

Building from source

Requires Bun 1.4+.

bun install
bun run dev            # run the CLI/TUI from source
bun run dev:desktop    # run the desktop app from source
bun run check          # lint + typecheck

Tests run from package directories, for example bun test inside packages/core.

Releasing the desktop app

Linux x86-64 only, distributed as an AppImage and a .deb through the Forgejo releases of this repository.

bun run build:desktop                          # build AppImage + .deb into packages/desktop/dist
bun run build:desktop -- --version=2.1.5-rc.1  # with an explicit version stamped in
bun run release:desktop v2.1.5                 # build + tag + Forgejo release + attach artifacts
bun run release:desktop -- --dry-run           # show what a release would do, no side effects

Releasing needs GIT_API_TOKEN in the environment or a .env at the repo root (scope write:repository); everything else — the Forgejo host and the x/harness repo — is derived from the origin remote. The release script is idempotent: it tags if needed, reuses an existing release, and replaces same-named assets, so re-running after a partial failure is safe. Release notes come from .release/vX.Y.Z.md when that file exists.

AppImage installs update in-app against the release feed (…/releases/download/latest/latest-linux.yml); .deb installs update by downloading the next release. Artifacts land in packages/desktop/dist — harness-desktop-linux-x86_64.AppImage and harness-desktop-linux-amd64.deb.

Syncing with upstream

bun run fork:status    # what changed upstream, which of our features it touches
git merge upstream/dev # then resolve per fork/features/*.yml instructions

See AGENTS.md and fork/README.md for the full workflow.

License

MIT, see LICENSE. Contains code from opencode by the opencode authors, licensed under the MIT license.