koigrid

Deploy apps and run managed Postgres, Redis, S3 storage, cron jobs and agent sandboxes on koigrid from any AI assistant (remote MCP with OAuth).

Hosted MCP Server

npx add-mcp 'https://koigrid.com/mcp'

Installs into Claude Code, Codex, Cursor and more

Documentation

Use koigrid from ChatGPT, Claude or any MCP client

koigrid has a remote MCP server. Connect it once and your assistant can deploy apps, create databases, read logs and check your spend — acting as you, with only the permissions you approve.

Server URL

Add this URL as a custom connector (remote MCP server) in your assistant:

https://koigrid.com/mcp

Connect

The first time, your assistant opens a koigrid window: sign in (or create an account) and approve what it may do and in which organization. No API key to copy. Your assistant gets a short-lived token, refreshed automatically.

Try it

Once connected, ask your assistant in plain words. For example:

  • “What do I have running on koigrid?” list_resources — your apps with status and URL, your databases and caches.
  • “Deploy https://github.com/me/my-app as a new app called my-app and tell me when it is live.” deploy_app, then get_app_status until the deploy is live or failed.
  • “My last deploy failed. Why?” get_app_status for the error, then get_app_logs with the build log.
  • “Create a Postgres database called my-db and connect it to my-app.” create_database, get_database_connection, then set_env_vars with a reference — the password never goes through the chat.
  • “How much have I spent this month? Cap it at $20 and warn me at 80%.” get_spend, then set_spend_limit (your assistant asks you to confirm).

Or use an API key

Clients that support a fixed header (Claude Code, Cursor and others) can send a koigrid API token instead of signing in:

claude mcp add --transport http koigrid https://koigrid.com/mcp --header "Authorization: Bearer koi_YOUR_KEY"

What it can do

A small set of task-shaped tools, each requiring its own permission. Read-only tools are marked so your assistant can run them without asking.

  • list_resources read-onlyapps:read Lists the apps (with status and public URL), databases and caches in the user's koigrid organization. Start here to find the id of an app or database before acting on it.
  • deploy_app apps:write · apps:deploy Deploys an app to koigrid. Pass appId to redeploy an existing app; or pass name plus a public Git repo URL (built automatically, Dockerfile or auto-detected) or a container image to create and deploy a new one. Pass env to deploy with its environment variables in the same deploy. Returns at once with the deployment id: the build runs in the background, so poll get_app_status until it is live or failed.
  • get_app_status read-onlyapps:read Shows an app's status, public URL, its latest deployments with their state and error, and how many times the running process died and was restarted in the last 24 h with the cause (restarts). Use it to follow a deploy, and when something fails read the error here first.
  • get_app_logs read-onlyapps:read Returns an app's logs: type "build" for the last deployment's build log (why a deploy failed), or "runtime" for what the running app prints (why it errors in production), from ALL its replicas and across restarts and deploys (since/until/replica/runner/q to narrow it down).
  • set_env_vars asks firstapps:write Sets environment variables on an app and redeploys it. Values are stored encrypted. To wire a koigrid database without pasting its password, use a reference such as ${{db.<database-name>.DATABASE_URL}}.
  • rollback_app asks firstapps:deploy Puts a previous deployment of an app live again, without rebuilding. Pass the deploymentId from get_app_status; without it, rolls back to the previous live deployment.
  • create_database databases:write Creates a managed Postgres database. Provisioning takes a few minutes: call get_database_connection afterwards to get the connection string once it is ready.
  • get_database_connection read-onlydatabases:read Returns the connection details of a database (host, port, user, database) and the reference to wire it into an app with set_env_vars. The password is never returned: the reference is resolved by koigrid at deploy time. Also reports the disaster-recovery copy (a replica in another provider that is never promoted automatically): null if the database has none.
  • set_network_access asks firstdatabases:write · redis:write Sets which IPs can connect to a database or a cache, like a security group: allowedIps null = public (any IP; password and TLS still required), a list of IPv4/CIDR = only those, [] = only from koigrid. The koigrid fleet (apps hosted here) is always allowed. Applied immediately; the result says if it was applied everywhere and warns when the IP of this request is not in the list. Existing connections are not cut.
  • get_migration_status read-onlydatabases:read For a database receiving a live migration (logical replication from an external Postgres): each subscription with its initial-copy progress (tables done of total), replication lag and a cutover verdict saying whether the app can be switched to koigrid yet, plus the remaining cutover steps. Never call the cutover done from lag alone: a subscription still copying tables can show zero lag. After the cutover (cutoverDone) it says which disabled subscriptions still hold a slot on the source that keeps growing its disk, and how to drop them.
  • get_database_queries read-onlydatabases:read Shows which SQL statements took the most execution time on a database in a time window (default: the last hour; hours up to 168, or since/until), with calls, total and mean time and their share of the database total, from pg_stat_statements sampled every ~2 min. Use it when a database or an app is slow, before guessing at indexes. Statement texts are normalized ($1, $2): no parameter values.
  • get_spend read-onlyusage:read · spend:read · billing:read Shows what the organization has reserved and spent this month, its spend limit if any, whether it has a card on file, its prepaid balance (wallet: balance, koigrid credit, automatic top-up, grace period) and the public price list (koigrid charges for reserved RAM and disk by the hour, shared vCPU and bandwidth included). To price something before creating it, POST /api/v1/pricing/quote needs no credentials.
  • set_spend_limit asks firstspend:write Sets a monthly spend limit for the organization, with optional alert thresholds and a hard stop that blocks creating new paid resources when it is reached.
  • create_vm sandboxes:write Creates a full Linux VM for an agent (sizes small 2 vCPU/4 GB, medium 4/8, large 8/16, xlarge 16/32), running and reachable by run_in_vm, and by SSH with the sshPublicKeys you give it (the result has the ssh command and the ~/.ssh/config block for VS Code / Cursor). It pauses on its own after autoPauseMin idle minutes (default 15, 0 = never) keeping its memory, and a paused VM pays no compute. Billed per second while running: medium $0.02/h (launch price; GET /pricing → agentVms for every size). Needs a card on file (add_payment_method).
  • run_in_vm asks firstsandboxes:write Runs a shell command in a VM (by id or name) and returns exit code, stdout and stderr (timeoutSec up to 600). If the VM is paused it is resumed first, with its memory and processes intact (resumed: true in the result); if it is still booting, the call waits up to 60 s. With detached: true it runs in the background with no time limit and returns a commandId at once (read it with manage_vm command_status).
  • manage_vm asks firstsandboxes:read · sandboxes:write Lifecycle of a VM (by id or name). action: pause (keeps memory, stops compute billing), resume, stop (keeps only the disk), start (cold boot from disk), fork (new VM that is an exact copy, needs name), snapshot (named, with memory, needs name), delete (removes the VM, its disk and snapshots: cannot be undone), or status. Omit vm to list the VMs of the project. Background commands: command_status (with commandId: its state and output from stdoutOffset/stderrOffset, up to 256 KB; without it: the list) and command_stop (SIGTERM to its process group, SIGKILL after 5 s). set_ssh_keys (sshPublicKeys) replaces the keys allowed to SSH in; the VM's ssh field has the command and the ~/.ssh/config block for VS Code / Cursor, and connecting wakes a paused VM.
  • add_payment_method billing:write Returns a link for the user to pay koigrid. With amountUsd (10-10000) it is a top-up of the prepaid balance: the user pays that amount on the provider page, gets an invoice, and the card is saved for automatic top-ups. Without amountUsd it only saves a card and charges nothing. The organization pays for the RAM and disk it reserves by the hour, shared vCPU and bandwidth included (prices: GET /api/v1/pricing). Use it when the free tier is not enough, the account is on the free-tier waitlist or the balance is low. Nothing is charged by this call: give the link to the user. billingMode chooses how reserved resources are paid: prepaid (default, each hour from the balance) or monthly (invoice to the saved card at the end of the month); pass it alone to only change the mode. Owner or admin only.
  • contact_support support:write Sends a message to the koigrid team: opens a support conversation, or replies in one when you pass its conversationId. Use it when something in koigrid fails and the error does not say how to fix it, when you are blocked, or when the user asks to talk to koigrid. Say what you tried and paste the exact error; pass appId or deploymentId if it is about one. A person answers; read the reply later with get_support_conversations.
  • get_support_conversations read-onlysupport:read Lists the user's support conversations with koigrid and how many have an unread reply from the team; pass a conversationId to read all its messages. Use it to see whether koigrid has answered.
  • mark_support_read support:write Marks a support conversation as read, after you have read the koigrid team's reply with get_support_conversations and told the user. It stops the unread-reply notice that koigrid adds to other tool results.

Security

Each connection appears in Settings → API tokens as a connected app. Revoking it disconnects it immediately, including its refresh token. Every action is in your audit log, and a connection limited to a project cannot touch resources of another project. Secrets never go through the chat: the assistant gets your database's host and user but not its password, and wires it into an app with a reference that koigrid resolves when it deploys.