For everyone deciding what to try next while waiting on a free GPU.
The execution environment
for AI agents.
AI agents like Claude and GPT access serious compute — including GPUs — over MCP. From running code to training jobs, with zero infrastructure to manage.
Copies Seyval's MCP server URL and opens the connector settings
Participating organizations
透明性・協力・標準化を基盤に、エージェンティックAIイノベーションの公共利益を推進するグローバルな中立オープン基盤組織。
Why Seyval
Your AI can design the experiment. It just can't run it.
It writes the code, but there's no GPU behind it. Matching the environment, submitting the job, watching the logs, collecting the results — that work still lands on you. An agent is only as capable as the things it can actually touch.
Give it an execution environment, and you can start asking for things like this.
"Register this repository and analyze the latest commit on main."
Manage the rest"Run train.py on an H100 and summarize the logs when it's done."
Run experiments"Sweep the learning rate over 1e-4 and 3e-4 — launch both."
Run experiments"Take the artifacts from that last run as input and just run the eval."
Run experiments"Read the failed run's logs and tell me if it's my code or the environment."
Run experiments"Register our lab's Slurm cluster and check whether 8 GPUs are free."
Manage the rest"Break down what each experiment cost this month by compute type."
Manage the rest"Turn this run's results into a share link I can send a collaborator."
Manage the rest"Analyze automatically whenever something lands on main."
Manage the restYou can start with a single repository.
How it works
From zero to GPUs for your agent, in minutes
Three steps to connect an execution environment to your AI agent. No DevOps, no Kubernetes manifests.
Link your repository
Just register the GitHub repo you want to run. Seyval analyzes the code at a given commit and surfaces the entry points and arguments you can run.
Your agent launches jobs
Agents like Claude and ChatGPT start runs over MCP. Just pick the compute you need, from T4 to H100.
Results come back automatically
Logs stream in real time and artifacts land in S3. The agent reads the results and moves to the next experiment.
Why Seyval
Serious compute for AI agents
GPU provisioning, job execution, logs, and artifacts — one platform your agents can drive autonomously.
Your agents
Seyval Platform
Cloud infra
T4 to H100, on demand
The right GPU is provisioned exactly when needed and scales to zero when done. No idle costs.
MCP-native by design
Built on the Model Context Protocol from day one. Claude or ChatGPT launches a job with a single tool call.
A disposable environment per run
Every run gets its own purpose-built image, and GPU jobs run one job per node. When the run ends the node is torn down, leaving nothing behind.
Durable execution
Long-running training and experiment jobs survive restarts, crashes, and deployments — Temporal under the hood.
Logs & artifacts, collected
Execution logs stream in real time, artifacts are saved to S3 automatically, and every run's cost is recorded.
Bring your own compute
Connect your Slurm/HPC cluster or lab machines and use them from agents through the same interface.
How this differs
Hand the GPU chores to your agent.
Rolling your own GPUs
Launch the instance, line up drivers and CUDA, install dependencies, and remember to shut it down. Every experiment repeats the same chore.
Pick a compute type. Image builds, node provisioning, and teardown are automatic, and the cost is recorded per run.
ML platforms built for people
Designed around a human submitting jobs from a console or SDK. Driving one from an agent means writing your own API glue anyway.
MCP-native from day one. Agents launch jobs with a tool call, then read logs and artifacts to decide the next move.
Driving HPC by hand
Write the sbatch script, convert the container to SIF, watch for free nodes, and relearn the conventions on every cluster.
Connect once over SSH and it behaves like the managed GPUs. Your code and data stay inside the cluster.
Use Cases
What teams run on Seyval
From autonomous research loops to model training — real workloads AI agents run on Seyval today.
AI-driven research loops
Agents iterate hypothesis → experiment → analysis on GPUs, autonomously. Researchers focus on direction, not infrastructure.
Learn moreModel training & fine-tuning
Hand off launching, monitoring, and retrying training jobs to an agent. From T4 to H100 — just point at your code and data.
Learn moreEnterprise ready
Built for scale, security, and control
Keep compute inside your own cluster
Connect your lab or corporate Slurm/HPC cluster over SSH and your code and data never leave your environment. Seyval only submits the job and collects the results.
Workspaces & access control
Three roles — owner, admin, member — govern access. SAML and OIDC federation with your corporate IdP is supported, configured per tenant at onboarding.
Every run on record
Which commit, on which machine, at what cost. Logs, artifacts, and spend are recorded per run.
Where your data lives
Everything stays in the Tokyo region (ap-northeast-1). Each region is isolated with its own infrastructure and KMS keys, and data is encrypted at rest and in transit.
Fits your existing workflow
Connect private repositories through the GitHub App and analyze automatically on push or PR. Works from any MCP-compatible AI client.