The platform for autoresearch
Give your LLM a supercomputer. Get thousands of GPUs on the only PaaS designed for agent-driven machine learning.
$ claude mcp add --transport http givemeanode https://mcp.autoresearch.sfcompute.comUsing Claude in the browser? Go to Settings, Connectors, Add custom connector and enter the remote MCP server URL: https://mcp.autoresearch.sfcompute.com
$ codex mcp add givemeanode --url https://mcp.autoresearch.sfcompute.comIn ~/.cursor/mcp.json:
{
"mcpServers": {
"givemeanode": { "url": "https://mcp.autoresearch.sfcompute.com" }
}
}In opencode.json (project or global):
{
"mcp": {
"givemeanode": {
"type": "remote",
"url": "https://mcp.autoresearch.sfcompute.com",
"enabled": true
}
}
}https://mcp.autoresearch.sfcompute.com$ devin mcp add givemeanode https://mcp.autoresearch.sfcompute.comOr in ~/.config/devin/config.json:
{
"mcpServers": {
"givemeanode": { "url": "https://mcp.autoresearch.sfcompute.com", "transport": "http" }
}
}In Capy, go to Settings, MCP servers, New MCP server, pick the HTTP transport, and enter the server URL:
https://mcp.autoresearch.sfcompute.comThen open the saved server and click Connect to finish the OAuth sign-in.
In ~/.config/amp/settings.json (or the Amp extension settings in VS Code):
{
"amp.mcpServers": {
"givemeanode": { "url": "https://mcp.autoresearch.sfcompute.com" }
}
}In ~/.omp/agent/mcp.json (or a project's .omp/mcp.json). /mcp add in a session does this for you:
{
"mcpServers": {
"givemeanode": { "type": "http", "url": "https://mcp.autoresearch.sfcompute.com" }
}
}In Zed's settings.json (zed: open settings from the command palette):
{
"context_servers": {
"givemeanode": { "url": "https://mcp.autoresearch.sfcompute.com" }
}
}Zed prompts for the OAuth sign-in on first use.
Then add credit. A new account starts with $0, and a node needs credit to start. Ask your agent to add $25, or run autor billing topup 25. With no card on file, either one answers with a payment link. One payment adds the credit and saves your card.
Make anyone a machine learning researcher
One sentence to your agent is the whole interface. The docs show every example as English, CLI, and the MCP calls underneath.
The full autoresearch suite
A full platform as a service designed and optimized for agents.
Interactive nodes
A container on an H100 with a persistent disk. Stop it and the files stay.
autor node create --chip h100 --wait
Batch jobs
Fire and forget. Queued time is free, checkpoints survive preemption.
autor job submit --chip h100 --context ./eval
Tasks
Thousands of work items, one container start. Lost machines cost only what was in flight.
autor job submit --tasks ./prompts.jsonl
Sweeps
Up to 256 variants of one job in a single call. One build, N runs.
autor api post /jobs/sweeps
Private networking
Your own nodes reaching each other over stable private addresses. Names resolve.
autor node create --network mesh1
Object storage
Org-owned buckets with a real S3 endpoint. Agents pull from them credential free.
autor storage cp ./train-data corpus/train/
Profiling
Locked clocks and unlocked profiler counters, set at create time and free.
autor node create --chip 8xh100 --profiling --clock-lock
Missions
Every job, log, and sample from one campaign gathered on a single page.
autor job submit --chip h100 --context ./train --mission trainer-v4
Metrics
Active series per month, first ten thousand free.
Logs
Logs are the interface. Follow a detached command to exit.
autor logs tulip -f
Traces
Spans ingested per GB, the first 5 GB each month free.
Samples
One small file per step, appended to a timeline you can scroll through.
autor api get '/samples?job=job-8x2mf'
RL sandboxes with reusable snapshots
Bake an environment once. Fork its saved state for each RL sample, with memory and running processes included.
Fork a snapshot, run each sample, and grade the results. Delete the other forks and snapshot the best result. Fork again from that snapshot.
Bake
Build the environment once and keep it as a snapshot. Baking bills as ordinary sandbox time.
autor sandbox bake --image sbx-task
Fork
One fork per sample, up to a thousand at once. Forking is free. 50ms per fork, 100ms per delete.
autor sandbox fork env-7fa2 --count 64
Snapshot and branch
Capture a running sandbox without stopping it, then fork the branch point. Idle sandboxes park themselves. 150ms per snapshot, tens of ms to wake.
autor sandbox snapshot sbx-0a1b
Sandboxes pause automatically while your agent is thinking, to save you 7.5x
Sandboxes park after 1.5s idle. Wake checks memory before work resumes. A 64-fork batch holding 2 GiB each across ten minutes of wall clock costs $0.98 held open the whole time and $0.13 with parking. Forks are free. The docs work through the arithmetic.
A new age of automated scientific discovery
Used by folks at
- Harvard
- MIT
- Stanford
- UC Berkeley
- UC San Diego
- UCLA
- University of Michigan
- UT Austin
“I LOVE givemeanode. I’ve been running astrophysics jobs on it. The agent-first design is amazing and the ticket experience has been remarkable.”
A fair queue nobody can buy out
We built a fair queue that lets folks buy a head start, but not the whole fleet. Everyone’s first node goes to the front - ahead of every paid tier - so you will always get rapid experimentation.
Least served goes next
100 jobs from one lab do not put you 101st. They dispatch one, then hold GPUs, so you dispatch second.
Always priority for the little guy
Your first node gets the biggest head start in the queue - bigger than anything money buys. Long live the experimenter.
A ceiling on the whole thing
Batch jobs never take more than 85% of the fleet, machines stay held back for wakes, and a returning owner is served before any new request at any price.
https://mcp.autoresearch.sfcompute.comAdd it to your agent, then ask the agent to add credit. A new account starts with $0, and the agent answers with a payment link. Then ask it for a GPU.