MCP-native · self-hostable · 20–500 people

One layer between everything your company knows and every AI that needs it

Connect your database with a read-only query, or any MCP server with its endpoint and two tool names. Contextely condenses each record into a small memory object, links it to the ones it relates to, and remembers where to go for the full text. Then every agent, assistant and script reads from that one place, over MCP, REST or the command line.

Start free Read the 5 minute quickstart

500 retrievals a month free, forever. Or run the whole thing on your own hardware with no limits.

Two connectors reach all of it: any MCP server, and a read-only SQL query.

Why this exists

Connecting AI to company knowledge is an N times M problem

Contextely turns it into N plus M. Connect each source once, connect each tool once, and the layer in the middle is the only thing that has to know about both.

Every tool needs its own wiring to every system

Five sources and five AI surfaces is twenty-five integrations, each with its own auth, its own sync job and its own idea of what a document is. Adding either a source or a tool multiplies the work rather than adding to it.

Each of them re-reads the whole company, every time

The default answer is to stuff a hundred documents into a prompt. That buys latency, a bill that scales with your document count, and a model that confidently averages three contradictory versions of the same fact.

Nobody can say what the assistant was allowed to see

The common design shows the model everything and asks it to redact. A model that has read a salary cannot be relied on not to allude to it, and there is no artefact proving it did not.

How it works

Three steps, and the third one never changes again

Connect what you already have

Point Contextely at a read-only SQL query against your Postgres, Supabase or Neon, or at any MCP server over Streamable HTTP: give it the URL and the names of a list tool and a fetch tool. Those two connectors are the whole surface, and both are configuration rather than an integration somebody has to build for your logo.

What connects today

It condenses, links and addresses

Each record becomes one small memory object: what is currently true, the entities and topics it is about, edges to the memories it relates to, and the address of the record it came from so the full text is one call away.

What a memory object holds

Everything reads from one place

MCP for agents that speak it, REST for everything else, a CLI for a terminal or a cron job. All three run the same retrieval, so what your assistant sees and what your script sees cannot diverge.

The quickstart

The middle of the diagram

A memory object is a fact, its neighbours, and the address of the original

This is what makes it cheap. Your agent gets three small objects instead of forty documents, and when three is not enough it follows the links or asks for the full record, on purpose, once.

memory objectfresh

Acme Industrial: renewal terms

Renews 2027-01-31 at $48,000/yr on the Growth plan. Owner Priya Raman. Negotiating a 14 month term for a 6% uplift. Two open commitments: SSO by 2026-10-01, audit-log export by 2026-11-15.

acme industrialpriya ramanrenewal

related memory

Q3 churn commitments to Acme
shares acme industrial, renewal
0.78
Support SLA by plan tier
shares growth plan
0.41

where the full record lives

source: crm_postgres
ref: accounts/acme-industrial
cite: crm.internal/accounts/acme
ttl: 900s
requires: sales

Illustrative object. Real shape, invented company.

The fact, condensed

One compact statement of what is currently true, written from the record by a model at ingest. This is what goes into your agent's prompt. Not the document, not forty chunks of it.

Links to related memory

Shared entities and topics make an edge. An agent that finds one useful memory walks to its neighbours instead of guessing at a second search query, which is the difference between one round trip and four.

The address of the original

The source, the record's stable reference in it, and a citation URL. When the summary is not enough, the agent asks Contextely to expand it and gets the whole current record back over the same connector. Nothing has to be stored to make that possible.

Every way in

The same retrieval, whichever end you hold it by

MCP, REST and the CLI are one code path with three doors. An agent, a backend service and a person in a terminal all get the same answer with the same entitlement and the same freshness labels on it.

MCP · agents that speak it

{
  "mcpServers": {
    "contextely": {
      "url": "https://www.contextely.com/api/mcp",
      "headers": { "Authorization": "Bearer ctx_sk_..." }
    }
  }
}

REST · everything else

curl -s https://www.contextely.com/api/v1/search \
  -H "Authorization: Bearer ctx_sk_..." \
  -H "Content-Type: application/json" \
  -d '{"query":"acme renewal","limit":5}'

CLI · a terminal or a cron job

curl -fsSL https://www.contextely.com/cli.mjs -o contextely && chmod +x contextely

./contextely config set-key ctx_sk_...
./contextely search "acme renewal"
./contextely related <memory-id>   # walk to what it relates to
./contextely expand <memory-id>    # read the whole record from the source

Identity

An API key resolves to a member, not to the workspace. An agent holding Priya's key sees exactly what Priya sees, so “the bot had broader access than the person who deployed it” cannot happen by construction.

Discoverable

An OpenAPI 3.1 document at /api/v1/openapi.json and an llms.txt, both generated from the same tool catalogue the MCP server advertises. They cannot describe an endpoint that does not exist.

The boundary

Who is asking is part of the ranking, not a filter on the answer

Every query carries an asker. Their entitlement scopes are applied as a factor inside the relevance score, so an object they may not see scores exactly zero and leaves the pipeline before a single character of it is read into an answer.

Because scoring runs first, an asker without the right scope never even causes Contextely to re-read that record from your database. That closes a timing side channel a redaction-based design leaves wide open.

src/lib/retrieval/score.ts

const entitlementFactor = entitled ? 1 : 0;

const score =
  relevance
  * (1 - W + W * freshness)
  * entitlementFactor;   // 0 never ranks

Read the security model

contextely.com/app

Ask the workspace

what are we committed to on the Acme renewal?

Ask as

Retrieve

asker priya@northwind.coconsidered 3returned 3withheld for entitlement 0refreshed 1

Acme Industrial: renewal terms
refreshed0.912
Renews 2027-01-31 at $48,000/yr on the Growth plan. Owner Priya Raman. Currently negotiating a 14 month term in exchange for a 6% uplift.
crm_postgresttl 900srequires sales
Engineering compensation bands 2026
fresh0.774
Band E4 midpoint 78,000 GBP, E5 midpoint 96,500 GBP. Reviewed 2026-07-02 by the compensation committee. Off-band offers need two approvals.
people_postgresttl 86400srequires people
Q3 churn commitments to Acme
fresh0.643
Two open commitments: SSO by 2026-10-01 and an audit-log export by 2026-11-15. Both tracked against the renewal above.
crm_postgresttl 900srequires sales

Illustrative data. Switch the asker to watch entitlement applied inside the ranking.

contextely.com/app/memory

Memory objects

5 of 100

ObjectSourceTTLState

Acme Industrial: renewal terms

fetched 4s ago

crm_postgres900srefreshed

Q3 churn commitments to Acme

fetched 3m ago

crm_postgres900sfresh

Support SLA by plan tier

fetched 2h ago

docs_mcp86400sfresh

Northwind pricing sheet v4

fetched 1h 12m ago

docs_mcp3600sstale

Engineering comp bands 2026

fetched 6h ago

people_postgres86400sfresh

One object is past its TTL. The next query that selects it re-reads the record from its source before it answers, rather than serving what is stored.

Never silently stale

Every memory knows how long it is allowed to be believed

A memory object carries the moment it was last read from its source and the TTL that source sets. When a query selects one that is past its window, Contextely re-fetches the underlying record over its connector, re-condenses it, and serves the current truth.

When the refresh cannot be done, because the source is down, the record was deleted, or the month's allowance is spent, the result comes back labelled stale_refresh_failed with the reason attached. There is no branch in this system where an out-of-date fact is served wearing a confident timestamp.

When the source has not changed, the raw payload hashes identically and the TTL clock resets without paying for a re-condensation, which is most of the bill on a real store.

Pricing

Priced on what costs us money, not on your headcount

Free is a real tier, not a demo. Self-hosting has no limits at all.

Free

$0/mo

Enough to wire it into a real system of record and prove it works.

1 connected system of record
100 condensed memory objects
500 retrievals / month across app, API and MCP
250 staleness refreshes / month
Entitlement enforced at retrieval time, by role
Full MCP server and REST API access
Self-host it yourself with no limits at all

Start free

Contextely Pro

$49/mo

Production volume, near-real-time freshness, and per-scope entitlement.

20 connected systems of record
10,000 condensed memory objects
25,000 retrievals / month
25,000 staleness refreshes / month
TTLs down to 60 seconds for near-real-time working memory
Per-scope entitlement grants, not just roles
Retrieval audit log: every query, every entitlement refusal
200 members

Start free, upgrade in-app

Full pricing detail, including what happens when you hit a limit, is on the pricing page.

Connect one source. Ask it a question. Five minutes.

No sales call, no pilot, no seat minimum. If it does not earn its place, the whole thing is a Docker image you can run yourself instead.

Start free Self-host instead