The context layer that checks itself.
MCP-delivered · free for individuals
The context layer that checks itself.
Stop re-explaining your data to your AI every session. Connect Claude Code, Cursor, Codex, or claude.ai in one command — your agent recalls your context before it writes SQL, and every definition is reconciled against your source, not just asserted. Bootstrapped from the SQL, dbt, notes, dictionaries, and semantic models you already have.
Claude CodeCursorCodexclaude.ai
Connect your AISee how it worksWatch the explainer · 71s
CLI connection example
$ claude mcp add --transport http clarilayer https://app.clarilayer.com/api/mcp/mcp --header "Authorization: Bearer cl\_YOUR\_CONTEXT\_KEY"
Your context layer
What your AI knows about your data
MCP connected
- Active customer Definition · you → reconciled Asserted
- Net revenue (excl. refunds) Definition · from dbt Asserted
- orders.region — null before 2023 Schema note · you → reconciled Caveat
- fct\_orders ↔ dim\_customers join
Reconciled against your source — not blindly asserted. Mismatches surface as a caveat so you and your agent know what to trust.
In the personal MCP path, ClariLayer never holds warehouse or CRM credentials, runs SQL, or calls HubSpot. Your agent uses its own source access and supplies the supported evidence; ClariLayer reconciles it and flags caveats.
The problem is measured — by the labs themselves
21% → 95%+
Analytics-agent accuracy before and after structured, checked context — “a context and verification problem, not code generation.”
Anthropic · 2026
95% → 65%
The same accuracy one month later, without maintenance. Context rots — something has to keep checking it.
Anthropic · 2026
< 1 pt
The gain from retrieval over their entire SQL corpus. More docs don't help — checked structure does.
Anthropic · 2026
Our own numbers, kept separate: in a pre-registered internal paired eval, the same agent scored 36/38 data questions with ClariLayer vs 26/38 without — and on the cheaper model, cost per correct answer flipped to $0.054 vs $0.081. The layer doesn't cost you money; skipping it does.
Also measured elsewhere — Gartner (2026): 60% of agentic-analytics projects relying on MCP alone will fail without a consistent semantic layer. dbt (2026): raw text-to-SQL reaches 64.5% — and fails silently.
Bootstraps SQL · dbt · CLAUDE.md · dictionaries · semantic models
See it work inside the agent
Your context, recalled and reconciled — without leaving Claude Code.
One command to connect. Then your agent recalls the definition you saved, runs the query with its own warehouse access, and reconcile flags what to treat with care.
claude-code — clarilayer
$ claude mcp add --transport http clarilayer https://app.clarilayer.com/api/mcp/mcp --header "Authorization: Bearer cl\_YOUR\_CONTEXT\_KEY"
agent
Connected. ClariLayer rides along inside the agent.
you
How many active customers do we have?
mcp
recall · use\_case: "active customers"
saved context · returned by recall
"Active customer" → fct\_customers,
status = 'active' AND mrr > 0
source: you → reconciled · status: asserted
agent
Running it with my own warehouse access…
SELECT count(\*) FROM fct\_customers
WHERE status = 'active' AND mrr > 0;
mcp
reconcile · declared vs. actual
!Caveat — declared definition excludes 142 rows where mrr is NULL; counted as not-active. Confirm NULL handling.
agent
You have 4,820 active customers.
Grounded in your saved definition and checked against your warehouse — with one caveat noted above.
Reenactment of the real loop — the same recall → reconcile → caveat flow the MCP runs today. Your agent keeps its own warehouse access; ClariLayer never holds your warehouse credentials and never runs SQL server-side.
Your AI is brilliant — and it forgets your data every session.
Every session starts from zero
New chat, new context window — and your agent has forgotten which table is the real one, how the joins go, and which definition you actually use. You re-explain it. Again.
A CLAUDE.md is just asserted text
Hand-typing your definitions into a notes file has the same trust problem as the original numbers: nothing checked it against your warehouse. It drifts, and your agent trusts it anyway.
So it repeats your data mistakes
Wrong table, wrong join, refunds counted in revenue, an active customer defined three different ways — the same mistakes, session after session, because nothing remembers the correction you already made.
Nobody's platform memory can tell you your revenue definition is stale — and every platform's memory is a silo. Remembering is easy. Checking is the hard part. It's the part we built.
Reconciled, not asserted
A CLAUDE.md is asserted. A ClariLayer entry is reconciled.
Hand-typing your definitions into a notes file is a real, useful start — but nothing checked it against your warehouse. A ClariLayer entry is the same definition, reconciled against your real warehouse result: provenance and status are shown, and a declared-vs-actual mismatch surfaces as a caveat instead of a confident wrong answer.
Hand-written CLAUDE.md
CLAUDE.md
\## Metrics
"Active customer" = a user who logged in
in the last 90 days (raw\_users.last\_login).
\# TODO: double-check this is still right
asserted · nothing checked it
Just text. Nothing reconciled it against your warehouse, so when the definition drifts the agent keeps trusting the stale line — confidently.
ClariLayer entry
"Active customer"
"Active customer" → fct\_customers,
status = 'active' AND mrr > 0
sourcedbt + you → reconciled
last reconciled2 days ago
statusasserted
caveat · declared ≠ actual
The declared definition counted logged-in users; reconciling against your warehouse showed the billing rule (`mrr > 0`) returns a different number. Flagged for you to resolve — not hidden behind a clean-looking total.
Reconciled against the result your agent computed locally. ClariLayer never holds your warehouse credentials and never runs SQL server-side — it checks declared against actual, and shows its work.
How it works
Bootstrap. Recall. Remember. Reconcile.
Four everyday verbs form the core loop. Supporting tools handle review, lifecycle, full-context fetches, and the completion receipt.
01
Bootstrap
Start from five shipped sources: SQL, dbt models, CLAUDE.md or notes, a data dictionary, and a semantic model. SQL, dictionaries, and semantic models become structured context; dbt and notes come along too.
02
Recall
Mid-task, your agent pulls relevant context in-flow. Exact names lead, use-case scoping stays explicit, and the result says when useful context was scoped away instead of silently hiding it.
03
Remember
Every correction, definition, and schema note you make persists. Your agent retains it across sessions, so you stop re-explaining the same thing every chat.
04
Reconcile
Ground a saved definition against a warehouse result or bounded, row-free HubSpot evidence. A mismatch surfaces as a caveat; otherwise the entry remains asserted.
Before the agent says done
Persist the agent's context declaration.
Before your agent says “done,” it files a receipt — what it learned, or an explicit nothing-to-update. Recorded durably, so the loop closes on the record instead of on vibes. What the receipt does and doesn't prove →
Same agent. Same prompt. The difference is your context.
Without your saved context, the agent guesses from raw table names. With it, the agent recalls the definition you reconciled against your warehouse — and knows what to treat with care.
Agent without your context
// You: "How many active customers do we have?" SELECT count(\*) FROM raw\_users WHERE last\_login > now() - interval '90 days'; // Wrong table (raw\_users, not fct\_customers). // "Active" = logged-in, not your billing rule. // The same mistake you corrected last week.
Agent with ClariLayer
// You: "How many active customers do we have?" // Agent recalls your saved context: "Active customer" → fct\_customers, status = 'active' AND mrr > 0 source: you → reconciled · status: asserted SELECT count(\*) FROM fct\_customers WHERE status = 'active' AND mrr > 0; // Grounded in your definition, checked against // your warehouse. Caveats flagged, not hidden.
The first reconcile
“Why don’t these two numbers match?”
The acute analyst moment: two sources disagree and the QBR is in an hour. ClariLayer reconciles a saved definition against your real warehouse result and flags the drift as a caveat — checked against source, never blindly asserted.
Your dashboard says
$1.42M
Net revenue for the quarter — the number on the board.
Your agent’s query returns
$1.51M
The ad-hoc query your agent just wrote, against the same warehouse.
A $90K gap, and no way to tell which one is right.
reconcile
Your agent runs the stored `net_revenue` SQL with its own access and reports the result shape back. ClariLayer compares the saved definition’s declared signals — its columns, grouping, and aggregates — against what came back.
Caveat recorded
They don’t line up: the live result is missing a column the saved definition expects, so reconcile records a caveat. That’s the signal to look closer — not a ruling on which dollar figure is right.
The caveat sends your agent digging
With its own warehouse access it traces the $90K to the cause: the ad-hoc query never joined the `refunds` table, so it reported gross, not net. You ship the refund-adjusted $1.42M.
remember
You correct it once so next quarter the agent starts from it instead of rediscovering it:
// saved to your context “net revenue must net refunds” source: you · status: asserted
The caveat caught the $90K discrepancy an hour before the QBR. Your agent traced it to the missing refunds join, you shipped the right number, and the correction is remembered for good. Your agent runs the query with its own access; ClariLayer never holds your warehouse credentials and never runs SQL server-side.
Live now · HubSpot CRM reconcile
Check the CRM rules hidden behind internal values.
Save a row-free contract for a HubSpot property — expected stages, a canonical value, deprecated aliases, or an expected label. Your agent reads HubSpot with its own authorized access and sends ClariLayer only bounded property metadata and value distributions.
On this personal path, ClariLayer never receives CRM credentials or records, and it never calls HubSpot. A mismatch becomes a `caveat`; otherwise the contract remains `asserted`. Salesforce contracts can be stored and recalled, but Salesforce reconciliation is not enabled.
See the HubSpot evidence recipe
Hand-run check · a design partner's live HubSpot
The obvious queries
Won dealsdealstage = 'closedwon'
MQLslifecyclestage = 'MQL'
What any agent would assumeThe labels the team uses every day.
What reconcile surfaced
Won deals caught20 of 2,401 · 0.8%
MQLs returned0 · 102 actual
OutcomeCaveat — internal values ≠ the team's labels
Hand-run with the partner's agent using its own access — bounded, row-free evidence only. The rules your CRM enforces are not the rules your team believes. Reconcile is how you find out before the QBR does.
Your context, working for you.
One analyst, your own data, your own agent. No governance rollout or procurement cycle — just your context, bootstrapped, reconciled, and remembered.
Day-1, not blank-slate
Bootstrap
Your agent doesn't start from zero — and it doesn't start from a hand-typed CLAUDE.md. Point ClariLayer at the work you already have.
What gets ingested
SQLValidated SELECTs, deterministically structured
dbtModel files imported as schema notes
CLAUDE.mdYour existing notes, brought along
DictionaryOne structured schema note per variable
Semantic modelOne canonical definition per model
“Bootstrap my context from ./analytics/sql + my dbt models.” Real working context on the first session.
Checked, not just claimed
Reconcile
A saved definition is reconciled against a warehouse result or bounded, row-free HubSpot evidence collected by your agent.
What reconcile returns
MatchApplicable checks pass; the entry remains asserted, not verified
MismatchFlagged as a caveat so you know what to trust
CRM boundaryNo HubSpot rows or credentials; ClariLayer never calls the provider
MeasuredInternal paired eval: 36/38 data questions right with the layer vs 26/38 without
Warehouse access and CRM access stay with your agent. ClariLayer compares only the supported evidence it sends.
It compounds
Remember
Every correction, every reconcile, every new note persists. Your agent grounds on more of your context over time.
The retention loop
Correct onceThe fix is remembered, not re-typed next session
Recall in-flowThe right context, pulled mid-task
CheckpointThe agent's updated or no-update declaration is recorded before completion
In the wildOne analyst found it cold, connected with zero onboarding, and ran 79 recalls across six active days — still active
“In normal use, we ask vague follow-up questions and expect ClariLayer to prevent us from re-explaining established business rules.” — SIRAYA, RevOps design partner. The context you build is the moat.
For teams · the Governed Context Edge
Yours today. Your team's — in pilot now.
ClariLayer starts as your personal context layer. The Governed Context Edge — your reconciled definitions promoted into a governed team canon, with ownership, approval, and one right metric — is built and in private pilot. A few design-partner slots are open.
Personal context merges into shared
The definitions and corrections you build solo are exactly what later merges into shared, owned team context — the same engine, one level up.
Conflicts adjudicated, not averaged
When two people's numbers conflict, the Edge surfaces it and a human settles it. Context proposes up and adopts down with agency — never a silent overwrite. One right answer, owned and approved, for everyone's agents.
Land solo, expand to the team
Start as your personal context layer with no procurement. When the team is ready, the Edge is already built and running in private pilot — design-partner slots are hand-onboarded, and a few are open now.
Request early accessSee the team story
Connect your AI
One command — or one OAuth approval.
Connect Claude Code, Cursor, or Codex with a context key, or add ClariLayer as a custom claude.ai Connector and approve over OAuth. Managed instructions keep supported local agents on the recall, confirmed-update, and completion-checkpoint loop.
Connect from claude.ai
OAuth · no context key
Add ClariLayer as a custom Connector, paste only the MCP URL, then sign in and approve in your browser.
https://app.clarilayer.com/api/mcp/mcp1. In claude.ai, open Settings → Connectors → Add custom connector. 2. Paste the ClariLayer MCP URL: https://app.clarilayer.com/api/mcp/mcp 3. Sign in to ClariLayer and approve the connection when prompted. 4. Back in a chat, confirm it worked by asking Claude to run the clarilayer\_\_health tool.
Connect a local coding agent
Claude Code, Cursor, and Codex authenticate with the context key you mint in ClariLayer.
Claude CodeCursorCodex
Run in your terminal
claude mcp add --transport http clarilayer https://app.clarilayer.com/api/mcp/mcp --header "Authorization: Bearer cl_YOUR_CONTEXT_KEY"Replace `cl_YOUR_CONTEXT_KEY` with your context key. The key is shown once at creation, so paste it straight into the command.
Built to be trusted with your data
Grounded against supported evidence — and it shows its work.
The whole point is that your AI's context is grounded against reality, not just asserted. We measured what that is worth — and we label every claim with exactly how far it goes.
Measured, not manufactured
36 of 38 with the layer. 26 of 38 without.
In our internal paired eval, the same agent answered 36 of 38 data questions correctly with ClariLayer connected — 26 of 38 without. On the cheaper model, cost per correct answer flipped: $0.054 with the layer vs $0.081 without. And under session load, the control went confidently wrong; the ClariLayer arm did not.
Shipped, not slideware
Four everyday verbs, with the supporting loop behind them.
Bootstrap, recall, remember, and reconcile are live in production, alongside review, lifecycle, full-entry fetch, and a durable completion receipt. Connect Claude Code, Cursor, or Codex with a context key, or claude.ai over OAuth.
A standard, not a shortcut
We built “verified” — and refused to ship it.
We built the stronger “verified” stamp, ran eight adversarial review rounds against it, and refused to ship it — one false “verified” is the failure a trust product cannot survive. Until a check can't lie, every mark says asserted or caveat, checked against a warehouse result or bounded, row-free HubSpot evidence. That refusal is the product.
Your data, your access
Personal reconciliation keeps source credentials with your agent.
On the personal MCP path, ClariLayer never runs SQL or calls HubSpot. Warehouse evidence can include optional preview rows; HubSpot evidence contains only bounded metadata and aggregate distributions, with CRM rows forbidden.
FAQ
Straight answers to the real objections.
What ClariLayer can see, what it doesn't touch, what's free, and what “reconciled” actually means.
Doesn't Claude or ChatGPT memory already do this?
Does my data leave my machine?
Does ClariLayer run my SQL?
Can ClariLayer reconcile HubSpot?
Is it really free?
Is my context “verified”?
What does a completion receipt prove?
What about my team?
Stop re-explaining your data.
Connect ClariLayer to Claude Code, Cursor, or Codex in one command. Free to start — no team, no procurement, just your context.
For teams
The Governed Context Edge is in private pilot — a shared, governed canon for your whole team's agents. We onboard design partners personally.
Work email
Company
Role
Request early access
We use privacy-friendly analytics
With your consent we use PostHog and Vercel Analytics to understand how ClariLayer is used so we can improve it. We never sell your data. Errors are always monitored (without analytics) so we can keep the app reliable. You can change your mind anytime.
DeclineAccept