MCP-native context infrastructure

Context infrastructure for your AI agents. In every tool. With provenance.

Connect your context once, and every agent can reach it — from Claude Code, Cursor, and ChatGPT to your own backend — over MCP and REST. One retrieval layer instead of a pipeline per tool, every fact traceable from commit to rank, at a fraction of the tokens.

One context layer across agent toolsA source and commit behind every answerFewer tokens spent finding the answer

No card required · first retrieval in ~5 minutes

quickstart

Two ways in. Five minutes to the first call.

Agents connect over MCP with OAuth handled for you. Your product calls the same retrieval over REST with an API key. Same context, same ranking, same provenance either way.

agent tools · mcp

Paste one endpoint into your MCP client.

Claude Code, Cursor, ChatGPT, or any MCP-aware agent. OAuth 2.0 with dynamic client registration means no keys to mint — your agent signs in and starts retrieving.

{
  "mcpServers": {
    "aidrop": {
      "url": "https://mcp.aidrop.it/mcp"
    }
  }
}
your product · rest

Search from your own backend.

One POST with tenant context. Every chunk returns rank signals, and include_provenance attaches the commit history behind each page.

curl https://api.aidrop.it/v1/context/search \
  -H "Authorization: Bearer $AIDROP_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "query": "how do we isolate tenant data?",
    "include_provenance": true
  }'

Quickstart, guides, Python and TypeScript SDKs, and the full API reference in the docs.

the problem

Your agents' context is stuck in silos you don't control.

Every platform ships retrieval that feeds only its own models, explains nothing, and still leaves you rebuilding the same plumbing around it.

01

Locked to one vendor.

OpenAI's file search serves OpenAI. Gemini's serves Gemini. Azure's lives in Azure. The moment your agents span tools — and they already do — your context doesn't follow them.

02

A black box when it matters.

When an agent answers wrong, nothing tells you which chunk came back or why it ranked. Opaque retrieval is the most repeated complaint developers make about platform search — debugging it is guesswork.

03

Plumbing you pay for twice.

Teams spend months on embedding pipelines and vector stores, then pay for idle clusters before the first query. The pipeline was never your edge — the context is.

portability

One context layer. Every tool your agents use.

Model vendors can't build this without feeding their competitors' models — an independent layer can. Describe your context once and stop re-integrating it per tool, per model, per quarter.

agent clients

Where your agents already work.

Claude Code, Cursor, ChatGPT, and any MCP-aware client hit the same retrieval endpoint. OAuth 2.0 with dynamic client registration is in production — connecting a new tool is a paste, not a project.

One MCP endpointOAuth + DCR built inNo per-tool pipeline
your product

Inside what you're building.

The same layer answers your backend over REST, with Python and TypeScript SDKs. Post documents, search with tenant context, ship the retrieval part of your AI feature in days.

REST API + SDKsTenant-scoped searchDays, not months
tomorrow's stack

Free to switch models.

37% of enterprises already run five or more models in production. Vendor retrieval can't follow a model swap — a context layer that belongs to you does, unchanged.

Model-agnosticTool-agnosticYours to keep
explainable retrieval

Every fact, traceable from commit to rank.

Every answer decomposes: which chunk, from which commit, recorded by whom, matched by which signal — full-text, semantic, or link-graph — and with what rerank score. In the dashboard and in every API response.

Your context is versioned like code, so provenance isn't a log we bolted on — it's the storage model. When an agent answers from your data, you can show your compliance team, your customers, and yourself exactly where every fact came from and how it changed over time.

Rank signals per chunkCommit history per pageCitations to the sectionNo other retrieval does this
architecture/tenant-isolation.md · § Row-level routingrerank 0.93
why this chunk · rank_signals
bm25
0.61
vector
0.84
graph
0.47
where it came from · commit history
3d2f9a1
Tighten OpenSearch routing per teammasha · 2 days ago · via mcp
8be04d1
Recorded from #eng-infra decisionmerge agent · Jun 24 · via signal
how it works

From your sources to a retrieval call you can defend.

Five stages, all managed — you never stand up a vector store, tune an index, or babysit freshness.

01 sources

Bring context, keep custody.

Connect a GitHub repo, Notion, or upload files. With git sources the content stays in your repository — we index chunks, not copies of your knowledge base.

02 ingest

Structure-aware ingestion.

Chunking follows your headings instead of arbitrary windows, with multilingual embeddings — your context works in English, German, or anything else your team writes.

03 index

Three-signal search.

Full-text, semantic, and link-graph retrieval run together, get fused, and a cross-encoder reranks the result. Most managed retrieval has no graph leg at all.

04 deliver

MCP and REST, same layer.

Agent clients connect over MCP with OAuth. Your product calls REST with an API key and SDKs. One retrieval layer behind both.

05 trust

Signals, history, audit.

Rank signals on every chunk, commit history on every page, and an audit trail attributing every call to a user and client.

the math

Cut retrieval cost. Here's the arithmetic.

To find or update context, an agent greps and re-reads — and the stateless API re-sends the whole growing transcript on every call. One targeted retrieval replaces the loop. Drag the sliders to your reality; the numbers update live.

20 steps
20
210×fewer input tokens with retrieval
Without aidrop.itgrep loop · transcript re-sent · tokens / month
2.8B
With aidrop.itone retrieval · the relevant slice · tokens / month
13M
Saved per month≈ with prompt caching · $5/1M input · illustrative
$2,838

How the estimate works: each grep/read step appends ~30k to the transcript, and the stateless API re-sends the whole transcript on every call — with 20 steps you pay for it roughly N(N+1)/2 times over. The re-sent transcript is billed at cache-read rates (~0.1× input) and each step's new output at cache-write (~1.25×), so the dollar figure already accounts for prompt caching. The alternative is a single ~30k-token retrieval — the relevant slice of your context, not the whole corpus — over 22 working days at $5/1M input.

See pricing for your usage
your plan limits

Spend your usage window on work, not re-reading.

Coding agents meter usage in windows — burn through yours and work stops until it resets. A grep loop bills the re-sent transcript against that window on every step; a retrieval bills once. At your slider settings that's 6.3M versus 30k tokens per task.

Without aidrop.itgrep loop · transcript re-sent
6.3M / task
With aidrop.itone retrieval · the relevant slice
30k / task
tokens billed against your usage window, per task — less is better
210×less of your usage window per task — the same plan covers that many more tasks before it resets.
proof, not promises

We run on our own retrieval — and measure it.

Our context-merge pipeline used to load the whole vault into the model on every run. We routed it through the same retrieval you get — fetching only the pages a write touches — and the cost fell off a cliff. Same output, a fraction of the tokens.

~6Mtokens per run — before
~300Ktokens per run — after
20×fewer tokens, faster & cheaper

Dogfooded daily

Our own context-merge pipeline runs through the same endpoint we sell. When it loaded the whole vault per run, we felt it in our bill — retrieval cut those writes 20×.

Measured, not vibes

A golden set of queries scores recall and ranking quality on every change, so search quality is a regression-tested number — not something we eyeball after a deploy.

Benchmarked in the open

We publish reproducible comparisons against black-box file search on realistic corpora. Want it proven on your own data? That's what the free tier is for.

teams & governance

Built for teams that answer to someone.

Security review shouldn't be where your retrieval vendor falls over. Isolation, access control, and attribution are in the foundation — because our first user was our own team.

isolation

Tenant isolation by default.

Every row carries the team, every index routes by it, every query filters by it. Multi-tenant separation is the storage model, not a setting you can forget to flip.

access

Roles and revocation.

Invite teammates with scoped roles. Remove someone and their OAuth tokens die with the membership — no orphaned access from a laptop you forgot about.

audit

Every call attributed.

Each retrieval and write is logged down to the user and the client that made it — the answer to "who asked what, from where" is a query, not an investigation.

custody

Your content stays in your git.

Connect your repository and the source of truth never moves — we index chunks for search, while your knowledge stays under your access control and your history.

for devtool teams

Your docs, inside your users' tools.

Your users don't live on your docs site — they live in Cursor and Claude Code. Connect your docs repository, hand them one MCP endpoint, and their agents answer from documentation that's current as of your last merge.

Not a chat widget on your website — your docs as a first-class tool inside the IDE where questions actually come up. And you see what your users ask, which is the roadmap input your docs team never had.

01Connect the docs repo — it stays your source of truth, synced on merge.
02Your users add one MCP endpoint to their agent, and it knows your product.
03You watch what they ask — unanswered questions become the docs backlog.
pricing

Priced like infrastructure: metered and predictable.

Usage you can compute before you commit — no surprise bills for idle clusters, no "contact sales" wall in front of a number.

Free tier, no card required
Pay for pages, queries, and storage
Every price calculable up front
See pricing
faq

Questions, answered.

If you don't see yours, email us - we read everything.

Yes — that's the point. One retrieval layer serves every MCP client (Claude Code, Cursor, ChatGPT and others) and your own backend over REST. Switch tools or models and nothing about your context changes.

Every chunk returns rank signals — how it scored on full-text, semantic, and graph search, and its rerank score. Ask for provenance and you also get the page's commit history: which commit produced the content, who recorded it, and when.

Both hit the same layer. Agent tools connect over MCP with OAuth handled for you; your product calls the REST API with an API key, with Python and TypeScript SDKs. Most teams end up using both.

No. Connect a GitHub repository and your content stays in it — we index chunks for search, and your repo remains the source of truth, synced as you merge. You can also bring Notion or upload files, all readable and deletable in your account.

No — vector search is one of three legs. Retrieval combines full-text, semantic, and link-graph search, fuses them, and reranks with a cross-encoder. You never operate the underlying stores, and no vector database gives you commit-level provenance.

Yes. Pricing is metered on pages, queries, and storage — computable before you commit, with a free tier that needs no card. No idle-cluster charges while you have zero traffic.

One context layer. Every tool.

Connect your context once — your agents retrieve it everywhere, with provenance for every fact.

Free to sign up · no waitlist · no credit card.