| name | massed-compute-ops |
| description | Operate Massed Compute GPU virtual machines through its MCP server. Use when the user mentions Massed Compute, massedcompute, MassedCompute, asks what GPU is available, wants to choose or launch training/inference VMs, inspect running instances, configure SSH access, audit hourly spend or billing runway, restart a VM, or terminate rented compute safely. |
Massed Compute operations
Use the Massed Compute MCP tools as the live source of truth for inventory, pricing, account state, and instances. Read references/recipes.md before the first account operation in a session.
Operating rules
- Validate the token before account work. If authentication is missing, request only
MC_TOKEN; never request the account password.
- Use live inventory and image compatibility data. Do not recommend a SKU or quote a price from memory.
- Before a launch, check token scope, billing settings, SSH keys, current capacity, price, compatible images, requested region, instance count, and naming.
- Treat launch and restart as cost-affecting reversible operations. State the chosen SKU, region, quantity, hourly cost, image, and name, then execute when within the user's request.
- Treat termination and SSH-key deletion as destructive. Show exact targets and costs, obtain explicit unambiguous user confirmation, call the tool, then re-list state to verify the target is gone.
- Never expose or persist VM passwords in chat, skill files, Git, or memory. MCP redacts them; the user can read them in the Massed Compute site if needed.
- After launch, poll instance state until usable, then provide the exact SSH target and record the instance UUID/name/GPU/region in the relevant project handoff or run journal.
- For long training jobs, preserve resumability: confirm durable storage/checkpoints, record shutdown criteria, and audit burn rate before leaving the job unattended.
Workflow router
- Inventory or GPU choice: call
gpu_inventory_list and images_list; rank two or three live options by VRAM fit, availability, total hourly cost, and architecture constraints.
- Launch: call
account_token_validation, account_billing, ssh_keys_list, gpu_inventory_list, and images_list before instances_launch; verify with instances_get or instances_list.
- Status: call
instances_list; use instances_get for exact details. Report UUID, name, GPU, region, state, age, and hourly cost.
- Spend audit: combine
instances_list, live product pricing, and account_billing; calculate per-instance and fleet hourly/daily burn plus recharge runway when balances are available.
- Restart: inspect the instance first, restart only the named UUID, then poll until healthy.
- Terminate: follow the destructive confirmation rule above; never terminate from a vague description without resolving it to exact UUIDs.
- SSH keys: list before creating. Add only a public key with a clear device name. Deletion requires exact disclosure, confirmation, and post-delete verification.
Tool availability
Read-only tokens expose inventory, images, instances, coupons, token validation, billing, and SSH-key listing. Full-access tokens additionally expose launch, restart, terminate, SSH-key creation, and SSH-key deletion. If a mutation tool is absent, identify the token scope as the cause instead of guessing.
Gotchas
- A 401 usually means the token was truncated; re-copy it unchanged.
- A 402 on launch means recharge billing is not configured.
- GPU/image compatibility is enforced; SXM-only and PCIe images are not interchangeable.
- A 429 is a rate limit; wait briefly and retry without duplicating a launch.
instances_list and instances_get redact cleartext VM passwords before the model sees them.- Environment-variable changes require a new Codex process before MCP authentication is available.
ssh_keys_create may return a generic upstream_error even for valid Ed25519 and RSA public keys. After two verified formats fail, stop retrying and have the user add the public key in the web settings; verify afterward with ssh_keys_list.- The live server may expose recipe discovery tools (
recipes_search, recipes_get, recipes_list) beyond the published lifecycle-tool table. Search recipes before provisioning software or a training stack.
Troubleshooting
- No Massed Compute tools: verify the MCP entry,
MC_TOKEN, network access, and restart Codex.
- Read tools exist but launch/terminate does not: replace the read-only API key with a full-access key only if mutations are required.
- Launch failed after a timeout: list instances before retrying; the first request may have succeeded.
- Cannot SSH: verify instance state, registered public key, username/image instructions, host/port, and local private-key availability.
- Cannot create an SSH key through MCP: validate the OpenSSH public-key format, try at most one alternative supported key type, then use the account UI and confirm the result through
ssh_keys_list.
- Unclear or stale price: refresh
gpu_inventory_list; never reuse an old quote.
