It reaches past one vendor
Any client that speaks MCP can be pointed at it with its own read-only credential. No per-vendor memory to migrate.
Free to start
What you told Claude never reaches ChatGPT — and neither has read your GitHub.
One context layer you own. Connect it once, and every AI tool you approve reads the same records.
One memory, shared across every AI you connect
Connect your sources once. Every AI tool you approve reads the same records.Provenance you can check, permissions you set, an export you can take.
Works with
- Claude
- ChatGPT
- Gemini
- OpenClaw
Claude · New conversationdemo
Write the release note for the auth change.
Auth moved to short-lived tokens two hours ago. That belongs in this release, not the last one.
GitHub · PR #4822h agoDemonstration. Nothing was pasted in.
Connect
Connect in two steps.
Mint a read-only credential
One per client, read-only, yours to revoke. Revocation is checked on every request.
- read-only
- one per client
- revocable
Point a client at the endpoint
JSON-RPC 2.0 over HTTP POST, MCP protocol 2025-06-18. This is the common shape.
{
"mcpServers": {
"superthinking": {
"url": "https://superthinking.ai/api/mcp",
"headers": {
"Authorization": "Bearer YOUR_KEY_HERE"
}
}
}
}It reads this workspace and cannot write, correct or delete. Authorization is a bearer credential — OAuth 2.1 is being built, and no specific client has been verified against the endpoint yet.
The difference
A week without it. A week with it.
Without SuperThinkingthree starts, three blank slates
One week · illustration
- Mon
You → assistant
Paste the architecture and the conventions. Again.
a blank slate - Tue
You → editor
Paste it again. It has never heard of you.
a blank slate - Thu
Browser chat
Proposes the approach you rejected in March.
a blank slate
With SuperThinkingone explanation, three reads
One week · illustration
- Mon
You → assistant
Explain it once. It becomes context you own.
your context - Tue
Your editor
Opens with that context. Nothing pasted.
your context - Thu
Browser chat
Knows the March decision, and cites it.
Notion · decisions
Say it once. Every approved tool reads it.
Start freeHow it works
Connect once. Read everywhere.
Step 01
Connect once
Authorize a source with a scoped grant. Widening what it reads later takes a new grant.
Step 02
It stays current
Connected sources keep feeding it, so what your tools read follows the work, not a file.
Step 03
Any approved tool reads it
A client asks over MCP and gets the slice that credential may see, source attached.
Step 04
You stay the owner
Provenance, scope, export and deletion are product surface. The ledger below names every limit.
Proof
One context. Every tool you approve.
17
source types you can connect
2025-06-18
MCP protocol revision served
JSON-RPC 2.0
over HTTP POST, one credential per client
Works with
- GitHub
- Linear
- Notion
- Slack
- Gmail
- Google Calendar
- Jira
- Sentry
- Vercel
- Stripe
- Intercom
- HubSpot
- PagerDuty
- RSS
- Forwarding inbox
- Signed webhook
Why this differs
Not a memory feature. Not a file. A layer you own.
Built-in memory only works inside the product that made it. A file you maintain freezes the moment you finish writing it.
It carries its age
Every item shows how old it is, so an ageing record is labelled ageing rather than served as current.
It is structured
Each item is its own record with a kind, an origin and a timestamp — not one long file you paste whole.
It leaves when you do
Context notes export as JSON Lines and delete for real, both written to the audit log. Where that is only partly true, the ledger says so.
Use cases
Four places you feel it.
A codebase across tools
Your editor and your browser chat start from the same picture.
Decisions that get relitigated
The approach you rejected in March arrives already ruled out.
Bringing someone new in
They get the context you scoped, and so do their tools.
Picking a project back up
Six weeks later the records are still there, with their age on them.
What is actually built
No invented numbers. Just the ledger.
Some of this runs and some of it does not. Here is which, including the limits on the parts that do.
Connect a source
live today17 source types today — repositories, trackers, inboxes, calendars, chat, monitoring, and a signed webhook.
Multi-tenant storage
live todayEvery record belongs to a workspace, and access is checked before anything is returned. Pushing that check into every query is still in progress.
Export your records
partly builtCSV runs today and is not gated by plan — but it returns the 500 most recent records, with no provenance column.
MCP endpoint
live todayJSON-RPC 2.0 at protocol revision 2025-06-18, authenticated per credential. Its tools are still the early-access set.
Context tools over MCP
in developmentSearch, fetch-with-provenance and explicit save, callable by any MCP client. Specified, not shipped.
Per-tool permission scopes
in developmentTwo scopes exist, and a read credential sees the whole workspace. Per-tool slices are being built.
Sign in with OAuth 2.1
in developmentAuthorization code with PKCE, discovery, refresh and revocation. Today a client authenticates with a scoped credential you create.
Delete everything, for real
in developmentYou can delete a source or a rule. Deleting the stored records behind it still goes through support.
No user counts, no uptime figures, no compliance badges. We hold no certification today.
Pricing
Start free. Pay when you outgrow it.
Two numbers separate the plans: the price, and how many records a month it covers.
More than a couple of sources? Start here.
Pro
For a workspace you work in daily.
$19/mo
No card required
- Monthly cap
- 10,000 processed records
Team
Same product, more headroom.
$99/mo
Enterprise is a conversation
- Monthly cap
- 50,000 processed records
On every plan
- Export as JSON Lines
- One read credential per client
- Workspace isolation
- Rate limits on the read path
The cap is the only plan difference the product enforces — what the tiers do not do.
Connect in two steps.
Back to setupFAQ
Straight answers.
What is SuperThinking?
A context layer you own. Connect a source once, and what it holds becomes context your approved tools read over MCP, with its origin and its age attached.
What actually changes day to day?
You stop re-explaining. The second tool starts from what the first one already learned, and can show you which record it came from.
How is this different from the memory built into my assistant?
That memory belongs to one product. You cannot inspect it, correct it, or take it with you. This is one layer every tool you approve reads, and it is yours to export or delete.
Why not just keep a context file by hand?
It freezes the moment you finish writing it, and grows too long to paste. Here the sources you connected keep feeding the layer, and every item carries its age.
How long does setup take?
Two steps: mint a read-only credential, then paste one config block into a client that speaks MCP. We have not measured that for anyone but ourselves, so we will not put a number on it.
What happens to my information?
We hold a derived view of the sources you connect, scoped to your workspace. We do not sell it and we do not train models on it. We hold no compliance certification today.