Docs, before there is a release
BAiD is not released, so there is nothing to install and no reference to read yet. This page documents what is already fixed: the idea, the interface an agent will use, and the endpoints this site serves today. The source is at https://github.com/froth-labs/BAiD.
The idea in four lines
- An agent alternates between thinking (a model call) and waiting (a lookup in its memory).
- The memory is usually a separate store: a graph database, a vector store, or files on object storage.
- BAiD sits between the agent and that store.
- Its job is to have the answer ready by the time the agent asks. The store is not replaced, migrated or modified.
Words we use
| Word | Meaning here |
|---|---|
| Agent | A program that runs a task as a series of model calls and lookups. |
| Memory | Whatever the agent looks things up in: a graph database, a vector store, object storage. |
| Think time | The time a model call takes. The agent is busy. |
| Wait | The time a lookup takes when the agent can do nothing else. This is what BAiD is built to remove. |
| Behind | Memory answers after the agent asks. How agents run today. |
| Ahead | Memory is ready before the agent asks. What BAiD is for. |
The interface (preview)
BAiD will reach agents as an MCP server with two tools. This is a preview. It is not available yet and may change.
| Tool | Arguments | Returns |
|---|---|---|
recall |
query (string), token_budget (integer) |
Context relevant to the query, sized to fit the token budget. |
remember |
fact (string) |
Confirmation that the fact was stored. |
A recall call:
{
"tool": "recall",
"arguments": {
"query": "what did this customer ask about last week?",
"token_budget": 1200
}
}A remember call:
{
"tool": "remember",
"arguments": {
"fact": "Customer prefers email over phone."
}
}Sample MCP client config, for when it ships:
{
"mcpServers": {
"baid": {
"command": "baid",
"args": ["mcp"]
}
}
}Do not add this config yet. There is no baid binary to run, and the command shown is a placeholder for the final one.
The waitlist API
This endpoint works today. It is how a person, or an agent acting for one, joins the waitlist.
POST https://frothlabs.in/api/subscribe
Content-Type: application/json| Field | Type | Required | Notes |
|---|---|---|---|
email |
string | yes | Your human’s email address. |
role |
string | no | For example “engineer”, “platform engineer” or “founder”. |
use_case |
string | no | One or two sentences on what their agent does and where its memory lives. Up to 500 characters. |
submitted_by |
string | no | Set to "agent" when you submit for a human. Defaults to "human". |
agent |
string | no | Your name, for example “Claude Code”. |
A successful request returns HTTP 200 and {"ok": true}. A bad request returns HTTP 400 and {"ok": false, "error": "..."}. A GET to the same address returns the field list as JSON.
What this site serves to agents
| Address | What it is |
|---|---|
/llms.txt |
Short summary and links to every page as Markdown. |
/llms-full.txt |
Every page in one file. |
/.well-known/agent.json |
Name, status, mode, links and contact as JSON. |
/<page>.md |
Any page as Markdown. Also returned for Accept: text/markdown. |
/sitemap.xml |
Every HTML and Markdown route. |
When BAiD is the wrong tool
- No think time. Lookups run back to back with no model call between them.
- Write-heavy workloads. Memory changes more often than it is read.
- In-process memory. The agent’s memory already lives inside its own process.
What comes at launch
Install steps, configuration, how BAiD works inside, and measurements with their method. None of that is published yet. The build log is where progress shows first.