Zektor

zektor

Command-line client for Zektor.io managed databases and caches.

Beta. The CLI and its MCP server are supported and live, but young. Expect

commands to be added and rough edges filed off as we learn how people use them.

Beta is not a warning about stability of what exists: **1.0.0 settles the

command surface.** Flags, output shapes and exit codes will not change without

a major version, so --json is safe to parse in a script, and tokens you

create now keep working.

Install

npm install -g zektor

Or run it without installing:

npx zektor whoami

Requires Node 18 or newer.

Authenticating

Create a token in the dashboard under Settings → Access tokens, then:

zektor login

The token is stored at ~/.config/zektor/config.json with mode 0600. It is verified against the API before being written, so a mistyped token fails here rather than confusingly on your next command.

--token exists for scripts, but prefer the prompt when you are at a keyboard — an argument lands in your shell history.

In CI

Set ZEKTOR_TOKEN and skip login entirely:

ZEKTOR_TOKEN=zk_… zektor whoami

ZEKTOR_TOKEN always beats the config file. zektor whoami reports which of the two supplied the credential, which is usually how you discover a stray environment variable is overriding the account you logged in as.

Commands

Every read command takes --json.

Trying it without an account

psql "$(npx zektor try)"

zektor try prints a connection string on stdout and, on stderr, how long it works and a claim link. The database accepts connections for 60 minutes. Open the claim link within 3 hours, sign up and add a payment method, and the data moves into a database of your own; otherwise it is deleted. Up to 200 MB and 3 connections, no superuser (trusted extensions such as pgcrypto work), and 3 trials per network per day. --json prints all of it as JSON. The same is available without the CLI:

curl -X POST https://api.zektor.io/api/trial

Creating

zektor db create --name=my-db --tier=AKPG-5 --region=nbg1 --engine=postgres@17

--tier, --region and --engine take the names you see in the dashboard and are resolved to ids for you. --tier and --region are required; leave one out or pass a wrong one and the error lists what is valid.

Creating runs in the background. Add --wait to block until the instance is ready; if creating it fails, the command prints why and exits non-zero.

--storage is rejected: storage starts at what the plan includes. Grow it afterwards with zektor storage resize.

Connecting

connect opens redis-cli for a cache. For Postgres it prints the command shape instead — the password is shown once when a role is created or rotated and the server never stores it, so there is no credential for the CLI to fetch. Create or rotate a role in the dashboard first.

Scaling and storage

zektor scale 42 --tier=AKPG-10

zektor storage show 42 zektor storage resize 42 --size=50 zektor storage autoscale 42 --on --limit=100

scale asks you to type the instance name, since it restarts the instance; --yes skips that. A downgrade warns that the smaller plan has less memory than the instance may be using. When the target plan includes less storage than the database is billed for, scale also says what the difference costs, with or without --yes.

scale and storage resize run in the background too, and take --wait: it blocks until the change has finished, and on failure prints the reason and exits non-zero. Without it, a failed scale is easy to miss, since the instance simply stays on its old plan.

Storage is PostgreSQL only — a cache is sized by its plan, so use scale.

• Included. Every plan includes storage: 5 GB on AKPG-5 to AKPG-40, then 10, 20 and 40 GB on AKPG-80, 160 and 320. A new database starts there.
• Growing. storage resize grows it any time, up to 1000 GB. It runs online, with no restart, and prints on stderr what the new size costs.
• Never shrinks. Storage only grows. Neither you, a plan change nor a restore can make it smaller. The only way to pay for less is to pg_dump into a new database and delete the old one.
• Price. €0.15/GB/month net above what the plan includes, in whole GB, prorated by the hour like plans. You pay for the size, not the usage: deleting rows frees room, not money. storage show prints what is billed now.
• Payment. Growing above the plan, or turning on automatic growth, needs a payment method or account credit.
• Automatic growth is opt-in, and needs a maximum above the current size: storage autoscale 42 --on --limit=100. At 80% full it grows to the larger of +5 GB or 70% full, never above the maximum. Each grow is permanent and billed like a manual one, and you're notified each time. After a failed automatic grow it waits a day. A limit can't be cleared once set, since the API reads an omitted value as "leave unchanged"; set a new number, or turn growth off with --off.
• Plan changes. An upgrade to a plan that includes more raises the size to the plan's, free. A downgrade keeps its storage, and the GB above the smaller plan are billed.
• Branches. A branch is billed on its plan, plus what it writes itself above its plan's included storage, while it is a branch. Once promoted, it is billed on its full size, which is the source's size when it was branched (or larger, if grown).
• Legacy databases. Databases created before growable storage can't grow yet. They are moving soon, and are never billed for storage.

Changed in 1.7.0. storage resize now grows storage through /api/instances/{id}/storage/grow and refuses a size that isn't larger than now. storage autoscale still accepts --min and --up-only, and ignores them. The MCP tool resize_storage now requires confirm_name.

MCP server

zektor mcp speaks the Model Context Protocol over stdio, so an editor or agent can list, create and scale instances directly. It uses the token you already logged in with — there is nothing extra to set up.

Add it to your MCP client's config:

{

"mcpServers": { "zektor": { "command": "npx", "args": ["-y", "zektor", "mcp"] } } }

For Claude Code, claude mcp add zektor -- npx -y zektor mcp does the same thing.

Tools

Guardrails

The CLI's protection against destroying the wrong thing is a prompt: it makes you type the instance name at a terminal. An MCP server has no terminal and its caller is a model, so that guard is translated rather than dropped.

• `delete_instance` is not registered unless `ZEKTOR_MCP_ALLOW_DESTRUCTIVE=1` is set. By default the tool does not appear in the list at all, so the worst a confused or injected instruction achieves is spending money, not losing data.
• `scale_instance`, `resize_storage` and `delete_instance` require a `confirm_name` argument that must match the instance's real name. It can only be filled in by having actually looked the instance up, which is the property the typed-name prompt had.
• Every tool carries the protocol's readOnlyHint / destructiveHint annotations, so a client that asks before running a tool asks about the right ones.

get_connection deserves its own note: for a cache it returns a connection string containing a live password, which means putting that password into the model's context. That is the point of the tool — it is how an agent wires an app up to a new cache — but it is worth knowing before you enable the server somewhere that logs conversations.

Diagnostics

The server writes its API URL, where it found a token, and whether deletion is enabled to stderr at startup; MCP clients collect that as server logs. A server pointed at the wrong API looks exactly like a broken one until you read them.

Output

stdout carries data and nothing else. Progress, warnings and errors go to stderr, so pipes stay clean:

zektor whoami --json | jq -r .email

Failures exit non-zero, so zektor … && next-command stops rather than continuing past an error the shell never saw.

Configuration

Roadmap

Next, roughly in order of how often they would be reached for:

• Backups — list, trigger, and restore to a point in time
• Roles — create and rotate Postgres credentials, which is also what would let connect work for Postgres rather than printing instructions
• Logs — tail an instance, the one thing people currently open a browser for

Not planned: admin commands. The admin console is deliberately browser-only.

Development

npm install

npm run gen:api # regenerate types/api.d.ts from the API's OpenAPI document npm run build node dist/index.js whoami

Point the whole thing at a local API with ZEKTOR_API_URL=http://localhost:5098, including the MCP server — ZEKTOR_API_URL=http://localhost:5098 node dist/index.js mcp serves your development backend rather than production.

npm run gen:api writes the API's full OpenAPI types to types/, which is gitignored — it describes every endpoint including the admin surface, and nothing imports it. It is a tool for checking the hand-written interfaces in src/api.ts against the real contract, not a build inpu