Skip to content
❋ ❋ ❋
The shared GPU compute platform

Idle GPUs, working together

Built on pooled compute. Powered by open-weight models.
Delivering batch AI at campus scale.

Submit a batch Lend a GPU Host it yourself
Runs open-weight models
Llama 3MistralQwen GemmaPhiDeepSeek
Platform

Everything a batch needs

From upload to download, the pool does the rest.

Batch API

Upload a JSONL of prompts, submit a batch, download results. OpenAI-compatible — existing code migrates by changing one line.

Distributed GPU workers

Jobs route to real GPUs pooled from labs and clusters. Workers pull work, so they run anywhere — even behind campus NAT.

Organizations & teams

A lab is an org. Owner, admin and viewer roles, shared worker keys, and every job attributable to the team that ran it.

Usage analytics

Requests, models, and per-worker contribution tracked from day one — the utilization evidence your institution wants.

Developers

One-line migration

BEFORE
base_url = "https://api.openai.com/v1"
AFTER
base_url = "https://your-deployment/v1"
api_key = "gk-your_personal_key"
# Everything else stays the same

OpenAI-compatible REST

Files and batches, verbatim. Your SDK already speaks it.

Model catalogue

Pick a pinned model id from /v1/models — copy, paste, run.

Dashboard

Submit, track, and download without touching curl.

Providers

Built for machines that come and go

Pull-based by design

Workers poll for work and disappear without ceremony. Intermittent lab machines are a feature, not a failure.

Fault tolerant

Silent workers are detected in minutes and their jobs requeued automatically. Batches survive the hardware they run on.

Curated model catalogue

Every servable model is a pinned artifact — weights, quantization, runtime. Reproducible by construction.

Put idle GPUs to work

It is software you host — a commons, not a cloud bill.

Read the self-host guide Work on the code

Home

Who are you?

Pick the one that describes why you are here. Each guide is meant to be read front to back, once. Everything assumes someone runs their own deployment — Sheshnag is software you host, not a service you sign up for.

I have prompts to run

Someone already runs a deployment and gave you a URL and a key. You swap base_url, submit a JSONL batch, poll it, download the results.

→ Run your prompts on Sheshnag

I have a GPU to lend

You want your machine to pick up jobs from a deployment someone else runs. One command, one key, about ten minutes — no repository clone, no Python, no database, no sudo.

→ Lend your GPU to Sheshnag

I want to change the code

Frontend, backend and daemon running locally, tests green, and the contract in front of you. It carries the process too — branching, review, who merges.

→ Work on Sheshnag

I want to run Sheshnag for my institution

You are standing up the control plane on your own premises: TLS, an admin account, the model catalogue, the first provider onboarded.

→ Run Sheshnag for your institution

Reference

Look these up; do not read them front to back.

Page What it answers
OpenAI compatibility Which OpenAI parameters are honoured, ignored, or rejected
Structured outputs Getting schema-constrained JSON out of a batch
Model catalogue How a model becomes servable — curation, pinning, digests
Data model Tables, relationships, and why they are shaped that way
Google OAuth Configuring Google sign-in for a deployment
Machine inspection What the daemon detects about a host, and how

Every reference page is dated and verified against the code it describes. Where a document and the code disagree, the code is the fact and the document is the bug — please open an issue.

Elsewhere in the repository

These live outside the documentation site because they address contributors and agents rather than product readers:

The component READMEs are stubs now: the API contract is API reference and the daemon's internals are Daemon internals.

Sheshnag is licensed under the Apache License 2.0 — see LICENSE.