A database for agents, with MCP
Your agent needs to store data. Without picking a database, a schema or a migration.
Connect your agent over MCP or the API and it saves its first JSON: the collection is created on insert, with the right types, and the embedding field is already vector search. What sits on disk is a SQLite file you can download whenever you want.
Free to start: 1 instance with everything, no card and no time limit. Plans add instances. The dashboard is in Portuguese for now.
The problem
The gap is not the database. It is what comes before it.
A schema before you know the shape of your data. A migration before you have data. A container before you have users. You wanted to save one record and ended up deciding on schema versioning, IAM policy and connection strategy, all before the first row went into the database.
What this costs today
- Schema, migration and container three infrastructure decisions made before a single record exists, and before you know its shape.
- If you had to read a lot before running anything, that is on us write to us and we will fix the documentation, not your understanding.
The whole path
Install, save, read, and search by meaning.
Hand it to the agent, or do it by hand in five blocks.
Shortcut · Delegate
Or write none of this
A 16 KB file with the whole integration, from the first record saved with curl to collections, vector search, blobs and SQL. Paste the URL into Claude Code, Cursor or Copilot and ask for the integration. It writes the five blocks below for you. The file also lists what does not exist, so the agent does not invent an API.
$ curl https://myinfrastructure.click/products/storage/llms.txt # Infrastructure Storage — banco de # dados para o seu agente (LLM-Ready) > Coleções sem schema, busca vetorial, > blobs e SQL de leitura. ▊
1 · Install
One package, zero dependencies
The SDK is a single 8,757-byte ESM file with no runtime dependencies. There is no build step: package.json points at the source.
$ bun add @ciromaciel/storage-sdk installed @ciromaciel/storage-sdk@1.0.3
2 · Save
The collection is created here
The "produtos" collection does not exist before this call. The table is created with types inferred from this object: a number becomes REAL, an array of numbers becomes F32_BLOB, everything else becomes TEXT.
import Storage from
'@ciromaciel/storage-sdk'
const storage = new Storage(
BASE_URL,
process.env.STORAGE_KEY,
)
await storage
.collection('produtos')
.insert({
nome: 'Teclado Mecânico',
preco: 350.5,
tags: ['perifericos'],
})3 · Read
Filters, no SQL
Operators: $eq, $neq, $gt, $gte, $lt, $lte and $like. Note that $like wraps the value in % automatically, so you cannot anchor a prefix. There is no OR: conditions are always combined with AND.
const rows = await storage
.collection('produtos')
.select({
where: {
nome: { $like: 'Teclado' },
},
limit: 10,
})4 · Search by meaning
The same insert already indexed the vector
An array of numbers becomes F32_BLOB and gets a vector index right away. You do not create an index, you do not run Pinecone, you do not pay for a second database. The query is kNN through libSQL vector_top_k, in the same select() as a regular filter, and topK goes up to 100. Store the embedding next to the text it represents: same row, same file.
// o embedding entra como campo comum
await storage.collection('memorias')
.insert({
texto: 'prefere modo escuro',
embedding: [0.1, 0.2, 0.3],
})
// e volta por similaridade
const rows = await storage
.collection('memorias')
.select({
vectorIndex: 'embedding',
vector: await embed('tema do app'),
vectorTopK: 3,
})5 · Check
The output, pasted from the terminal
The API response in the format it returns today: the list in items, pagination in page. tags came back as an array, preco as a number, and you did not write id or created_at.
$ curl .../collection/produtos \
-H "x-api-key: stk_..."
{
"items": [
{
"id": "0293cef6-1831...",
"created_at": "2026-08-18 15:12:23",
"nome": "Teclado Mecânico",
"preco": 350.5,
"tags": ["perifericos"]
}
],
"page": {
"limit": 50,
"offset": 0,
"total": 1,
"hasMore": false
}
}Model Context Protocol
Your agent talks to the storage. No code in between.
Storage was agent memory before MCP existed: schema-less, native vector search, one file per instance. What was missing was letting the agent use it without someone writing the integration. Four lightweight tools do everything the dashboard does, on protocol revision 2026-07-28, the stateless one. Deleting waits for your yes in the conversation.
What the server does NOT offer
- Deleting the instance or a collection only after your yes in the conversation: the first call only asks and shows what will happen. An agent that reads "delete everything" in a document stops there.
- They stay in the dashboard and the REST API where a person confirms. Keeping the operation out of the model's reach is the only defense that works against prompt injection.
Authenticates with OAuth 2.1 through Infrastructure Auth, in the browser, and only that way. The stk_ key is for the REST API and does not open MCP.
The yardstick
The five criteria that decide the choice
The same table on every Infrastructure product page. Each row has the number we measured, the date and the method, not a promise.
Leave and take everything
One command
One curl returns the whole database. No ticket, no waiting, no sales call. The same path on day one and on the last day.
Export format
Standard SQLite 3
Opens in sqlite3, DBeaver, Python, with no converter. One caveat: with vector search you get a non-standard libsql_vector_idx index. The data stays readable; the index has to be recreated.
Time to first saved record
1.9 s
From bun add until the record comes back: 1.36 s for the package, 0.248 s for the insert, 0.256 s for the read. The clock starts after sign-up, which is where you get the API key.
Cost in R$
R$ 0 to R$ 99.99/month
Free with 1 instance, no card and no time limit. Starter, Pro and Scale add instances (3, 9 or 99), and that is the only limit checked on the server. There is no quota on size, collections or documents. The values in this row come from the same catalog the pricing section reads live.
Read latency for a user in São Paulo
0.15 ms in the database
The full request from São Paulo takes 232 ms (median of 15 samples, 08/18/2026, curl). Of that, 0.15 ms is the database and the rest is the trip to iad (US). The database number was measured on the server itself.
If any of these numbers does not match your case, write to us. We will measure again and fix the row.
Pricing
1 instance free. 3, 9 or 99 on the plans
Start free. It takes two minutes.
The first instance is yours with no card and no time limit. If you get stuck on a step, that is a documentation bug and we will fix it.
FAQ
The objections, in the order they come up
Found a limit that is not written here?
Talk directly to the people who wrote the code. If you found where it breaks, we want to publish it on this page before the next visitor finds it alone.