Riven Computer — User Guide

Getting Started

1. Navigate to rivenai.io/computer
2. Type a task or question in the input bar
3. Watch as Computer plans, searches, reads sources, executes code, and synthesizes an answer
4. Download any generated artifacts (CSV, JSON, reports)

Web Search

Computer searches the web in real-time using the Riven search index. Every response includes cited sources with domain chips you can click to visit the original page. Sources appear both inline in the response and as clickable chips below each answer.

URL Fetching

Computer automatically reads the top search result to extract deeper content beyond the snippet. This gives you full-page context, not just search summaries. The fetched content is displayed in the step trace and included in the final synthesis for more accurate, detailed answers.

Code Execution

Computer generates and executes JavaScript code in an isolated sandbox environment. This enables:

• Calculations — Fuel costs, ROI, pricing models, financial projections
• Data processing — CSV/JSON generation, data transformation, sorting, filtering
• File generation — Downloadable artifacts in CSV, JSON, markdown, HTML, or SVG format
• Analysis — Statistical computations, comparisons, formatting

Each code execution shows the generated code, terminal output (stdout/stderr), execution duration, and any produced files as downloadable artifacts.

Artifacts

Generated files appear as clickable cards below the response. Click any artifact to preview the file content in the side panel, download the file to your device, and see file type, size, and container metadata. Artifacts persist across multi-turn sessions within the same sandbox.

Multi-Turn Sessions

Each conversation uses a persistent sandbox. When you send a follow-up message, the same sandbox container is reused, files from previous turns are accessible, conversation history provides context, and prior execution count is displayed in the sidebar. Sessions now persist to disk, so they survive server restarts and redeploys — your sandbox, files, and history are restored automatically.

File Uploads

Click the paperclip icon in the input bar to upload files to your sandbox. Supported types: CSV, JSON, TXT, Markdown, PDF, PNG, JPEG, WebP, GIF, XLSX. Files are stored in the sandbox uploads directory and are accessible to the agent during code execution. The agent automatically detects uploaded files and can read, analyze, and process them. Maximum file size: 10MB.

Scheduled Tasks

Create recurring tasks that run automatically on a schedule. Click the + button next to “Scheduled” in the right sidebar to create a task with a prompt and interval. Available intervals: 5m, 15m, 1h, 6h, 12h, daily, weekly. Tasks persist to disk and survive server restarts. Each task tracks run count and stores the last 10 results. Pause, resume, or delete tasks from the sidebar.

GLM Agent Flights

A flight is a coordinated group of GLM agents that work on related sub-tasks in parallel. Click the lightning bolt button in the right sidebar to open the flight dialog. Add agents with names and instructions, then spawn the flight.

Each agent in a flight gets:

• Isolated sub-sandbox — distinct working directory and file namespace
• On-prem model routing — routes through on-prem GPU models first, falls back to cloud
• Provenance tracking — each agent reports its model, route used (onprem/gateway), and sandbox ID
• Automatic synthesis — results from all agents are combined into a single structured report

The on-prem model (Qwen 3.6 35B on NVIDIA A100) serves as the primary route. If the on-prem model is unavailable or rate-limited, the flight automatically falls back to cloud models through the Riven gateway. The route used is visible in each agent's status response.

Browser Tool

Click the globe icon in the input bar to fetch and extract content from any URL. The browser tool sends a server-side request, strips HTML, and returns readable text content in a dialog with the page title, URL, and scrollable content view. SSRF protection blocks local/internal addresses. 15-second timeout per request.

Step Trace

Every response includes a collapsible step trace showing:

• Planning — How Computer analyzed your task
• Search — What queries ran and results found
• URL Fetch — Pages read for deeper content
• Code Execution — Generated code, terminal output, duration
• Synthesis — Final answer construction with streaming
• Completion — Summary with source count, execution count, duration

Click “Completed N steps” to expand or collapse the full trace.

Usage & Credits

Each task consumes credits based on the operations performed:

Credit usage is displayed in the right sidebar. Typical tasks cost 50-100 credits.

Use Cases

Safety

Code execution includes safety guards that block:

• Environment variable access (prevents credential leakage)
• Process spawning (sandbox isolation)
• Path traversal attacks
• Infinite loops
• Raw network requests

If code is blocked, the step trace shows the specific safety rule that triggered.

Model Routing Architecture

Riven Computer uses an on-prem-first routing strategy for all agent and flight executions:

1. On-prem primary — Qwen 3.6 35B served via vLLM on NVIDIA A100 80GB at 172.30.4.1:11515. Zero network latency, no per-token cost.
2. Cloud gateway fallback — GLM-4.5-flash via Fireworks through api.rivenai.io. Activated automatically when on-prem is unreachable or rate-limited.
3. Provenance in every response — routeUsed field shows which route served each request (onprem or gateway).

A separate model selector next to the mode toggle picks which on-prem or cloud model answers, independent of the fallback chain above:

1. Quick — Qwen 3.6 35B, on-prem (172.30.4.1:11515). Default. Fast, simple queries, 180-token cap.
2. Deep — Nemotron 3 Puzzle 75B, on-prem (172.30.4.1:11530). Complex or research queries, 1500-token cap.
3. Fast — riven-instant via the cloud gateway. Opt-in only — on-prem is never bypassed automatically.
4. Auto — classifies the query and routes to Quick or Deep. Never routes to cloud on its own.

Each agent in a multi-agent flight gets its own isolated sub-sandbox with a unique working directory. Sandboxes are tracked in the runner service and visible via the health endpoint.

FAQ