Point any AI at any of your knowledge bases.

MCP, the agent memory surface

Four tools. Read + add only. Update / delete / organize live on the REST surface below.

REST API, the management surface

Beta

Full CRUD over folders, items, and memberships. Built for automations + integrations (n8n, Zapier, cron scripts). All endpoints described in the /api/v1/openapi.json spec. Drops into any OpenAPI-aware tool wrapper.

Endpoint shapes may change as we learn from real usage. The MCP server is the more stable surface today — build production integrations there where you can; reach for REST when MCP can’t cover the workflow.

Scope a token down

Every token carries three scope dials. Use them to give an AI client exactly the slice of the library it needs and nothing else, the way you’d give a Unix process the smallest set of file permissions that gets the job done.

Curl

# Search the user's knowledge base
curl -X POST https://graniite.co/api/mcp \
  -H "Authorization: Bearer lev_..." \
  -H "Content-Type: application/json" \
  -d '{
    "jsonrpc": "2.0",
    "id": 1,
    "method": "tools/call",
    "params": {
      "name": "search_knowledge_base",
      "arguments": { "query": "what did that podcast say about agents" }
    }
  }'

# Add a URL (requires write capability)
curl -X POST https://graniite.co/api/v1/ingest \
  -H "Authorization: Bearer lev_..." \
  -H "Content-Type: application/json" \
  -d '{
    "kind": "url",
    "url": "https://www.youtube.com/watch?v=..."
  }'

Vercel AI SDK

Wire Graniite directly as an AI SDK tool. The MCP server runs over HTTP, so the standard experimental_createMCPClient works with no adapter package needed.

import { experimental_createMCPClient as createMCPClient, generateText } from "ai";
import { gateway } from "@ai-sdk/gateway";

const mcp = await createMCPClient({
  transport: {
    type: "sse",
    url: "https://graniite.co/api/mcp",
    headers: { Authorization: `Bearer ${process.env.GRANIITE_TOKEN!}` },
  },
});

const { text } = await generateText({
  model: gateway("anthropic/claude-sonnet-4-6"),
  tools: await mcp.tools(),
  prompt: "Summarize what's in my Graniite library about LLM observability.",
});

await mcp.close();

Hybrid search

search_knowledge_base accepts a mode parameter. Default is hybrid (vector similarity blended with Postgres full-text rank), which catches both fuzzy semantic matches and exact strings (proper nouns, version numbers, library names) that pure vector search smooths over. Override to semantic or keyword when you have a reason to.

{
  "name": "search_knowledge_base",
  "arguments": {
    "query": "Postgres logical replication",
    "mode": "hybrid",
    "limit": 8
  }
}

Upload binary content

For audio, video, PDF, or images, request a signed URL, PUT the bytes to it, then call finalize. The pattern bypasses the serverless body-size cap and lets agents drop files up to 50 MB straight into the user’s library. Audio and video go to /finalize; PDFs and images go to /finalize-document. The signed-URL response tells you which.

# 1. Ask for a signed URL
curl -X POST https://graniite.co/api/v1/uploads/signed-url \
  -H "Authorization: Bearer lev_..." \
  -H "Content-Type: application/json" \
  -d '{"mime":"audio/mpeg","size":4823104,"filename":"meeting.mp3"}'
# → {"path":"...","signed_url":"...","finalize_endpoint":"finalize"}

# 2. PUT the bytes directly to signed_url
curl -X PUT "<signed_url from step 1>" \
  -H "Content-Type: audio/mpeg" \
  --data-binary @meeting.mp3

# 3. Finalize → ingestion runs server-side
curl -X POST https://graniite.co/api/v1/uploads/finalize \
  -H "Authorization: Bearer lev_..." \
  -H "Content-Type: application/json" \
  -d '{"path":"<path from step 1>","mime":"audio/mpeg","size":4823104,"filename":"meeting.mp3"}'

OpenAPI spec

The REST surface (/api/v1/ingest today, more endpoints later) ships an OpenAPI 3.1 spec. LangChain’s OpenAPIToolkit, LlamaIndex’s OpenAPIToolSpec, and most agent frameworks can consume it directly without us writing per-framework adapters.

curl https://graniite.co/api/v1/openapi.json

What it costs you

Read calls (search_knowledge_base, get_item, list_recent) are free on every tier. Write calls (add_to_knowledge, /api/v1/ingest) for paid-transcription content (audio, video, YouTube) count against your monthly transcription quota, the same quota the UI uses. Webpage scrapes and pasted text are free. Quota state is visible at /settings; quota errors return HTTP 402 with the tier and cap so your client can render an upgrade prompt.

MCP + REST API access is a paid-tier feature. Free users get the in-app experience but can’t mint tokens. Token-create calls return HTTP 402 with {tier, cap, used} in the body. See /pricing for tier shape; above the gate, every paid tier issues unlimited tokens.

Per-token visibility

Every call logs to /settings/integrations so you can see what your agents are doing: query strings, latency, result counts, which token. 90-day retention, clearable from the UI.