Skip to content

opencode-assistant

Directory actions

More options

Directory actions

More options

Latest commit

 

History

History

opencode-assistant

Folders and files

NameName
Last commit message
Last commit date

parent directory

..
src
src
 
 
.gitignore
.gitignore
 
 
README.md
README.md
 
 
agentuity.json
agentuity.json
 
 
app.ts
app.ts
 
 
bun.lock
bun.lock
 
 
package.json
package.json
 
 
tsconfig.json
tsconfig.json
 
 
vite.config.ts
vite.config.ts
 
 

README.md

OpenCode Assistant

Demonstrates Agentuity platform features using a chat interface for GitHub repositories:

  • Sandboxes: Create, configure, and destroy isolated runtime environments
  • KV Storage: Persist workspace state across stateless HTTP requests
  • SSE Streaming: Proxy server-sent events from a sandbox to the browser
  • End-to-end Type Safety: Valibot schemas flow from API routes to React hooks
  • AI Gateway: Route LLM calls through Agentuity's gateway with a single SDK key

No agents are defined. All features are used directly from API routes.

Note: This example requires live sandbox infrastructure via Agentuity Cloud. The bootstrap process creates a remote sandbox, installs OpenCode, and clones a repository, which can take 30-60 seconds. Transient failures during sandbox provisioning or OpenCode startup may cause health check timeouts. If you run into issues, retry or reach out in our Discord help channel.

Getting Started

bun install
bun run dev

No separate provider API key needed. The AGENTUITY_SDK_KEY (set automatically by agentuity auth login) routes LLM calls through the Agentuity AI Gateway.

Open localhost:3500 for the frontend, or localhost:3500/workbench to test the API directly.

How It Works

  1. Enter a GitHub repo URL (defaults to https://github.com/agentuity/sdk)
  2. Click Start: a sandbox boots with opencode:latest, clones the repo, and starts the OpenCode server
  3. Ask questions about the codebase: responses stream back via SSE with markdown rendering
  4. Click Stop to destroy the sandbox

The backend creates the sandbox without a command (which returns immediately), then runs setup in a fire-and-forget execute call:

// Create sandbox without command — returns in ~2s
const sbx = await sandbox.create({
  runtime: 'opencode:latest',
  network: { enabled: true, port: 4096 },
  resources: { memory: '2Gi', cpu: '1000m' },
  timeout: { idle: '30m' },
  env,
});

// Fire-and-forget: clone repo + start OpenCode with a watchdog loop
sbx.execute({
  command: ['bash', '-c', [
    'mkdir -p ~/.config/opencode',
    "echo '{\"model\":\"anthropic/claude-sonnet-4-6\"}' > ~/.config/opencode/opencode.json",
    'git clone --depth 1 $REPO_URL ~/project || true',
    'mkdir -p ~/project',
    "nohup bash -c 'while true; do cd ~/project && opencode serve --port 4096 --hostname 0.0.0.0 >> /tmp/opencode.log 2>&1; sleep 2; done' > /dev/null 2>&1 &",
  ].join('\n')],
}).catch((err) => logger.warn('execute() rejected', { error: String(err) }));

Session Persistence

This example keeps one active workspace at a time. KV stores the current sandbox ID, server URL, repo URL, and OpenCode session ID with a 30-minute TTL. Refreshing the page reconnects to that same running workspace, and starting the same repo again reuses it instead of re-cloning.

If you start a different repo, the existing workspace is torn down and a fresh sandbox is created for the new repo. The OpenCode session persists only for the currently attached workspace, so the AI retains prior context across page refreshes but not across repo switches. The visual chat history still resets because it lives in React state.

Project Structure

src/
├── api/index.ts        # API routes: start, status, ask, events (SSE), stop
├── lib/
│   ├── opencode.ts     # OpenCode HTTP helpers (health, session, prompt)
│   └── types.ts        # Shared schemas (AssistantState, request/response types)
└── web/
    ├── App.tsx          # Chat UI with repo input, markdown streaming, stop controls
    └── App.css          # Agentuity brand theme (cyan + zinc)

Related