ClariLayer

ClariLayer keeps the facts, preferences, decisions, rules and lessons you choose to carry forward. Connect Claude Code, Cursor, Codex, claude.ai or another compatible remote MCP client. Your AI can recall relevant work context from the selected authorized space, with its source and applicability visible.

Start free. No credit card for the current personal offer.

One decision, carried into the next session

You settle a launch decision with your AI:

For the Launch project, show the product demo first. Technical detail belongs in docs. Remember this confirmed decision with its source and project applicability.

In a later session:

Recall the Launch project's saved decisions before drafting the launch page. Show which decision applies and where it came from.

The decision can travel across projects and sessions within the authorized space while retaining its original applicability. A rule for one customer or project stays bounded. If it changes, correct it; if it no longer belongs, use scoped forget or exclusion.

Recall depends on the AI client actually calling the tool. Connecting MCP does not guarantee every answer uses saved context.

Related MCP server: CacheTank MCP Server

Connect your AI

Use Connect your AI for guided setup. Select your client and space, give the setup request to that AI, and complete authentication in the client. Use OAuth when the selected route supports it, or store a context key locally as the fallback. Never paste a key into an AI conversation or a GitHub issue.

Already connected? Choose Update my AI setup on the same page, then run its request in the intended project. This updates one recognized ClariLayer instruction block while preserving your own instructions.

The installed bootstrap obtains the current workflow from get_project_stanza with mode: "runtime" once at the first relevant use in each new AI session. A routine server guidance update then reaches the next session without copying another long workflow into every project.

Optional setup CLI

The npm CLI supports context-key connection setup and a local dbt check. CLI 0.3.0 replaces the embedded Analytics instructions with a request for your connected AI and adds instructions --agent <client>.

Check which version is available before choosing a command:

npm view clarilayer version

With 0.3.0 or later available:

npx clarilayer@latest init
# Already connected? Print only the instruction update request:
npx clarilayer@latest instructions --agent codex

If your registry still serves 0.2.1, use npx clarilayer@0.2.1 init --no-stanza to skip its older embedded instructions, then complete Update my AI setup through the guided setup above. The hosted path is available independently of npm publication.

See CLI.md for options and the source-checkout route, CHANGELOG.md for release history, and RELEASING.md for publication checks. Native instruction qualification records the tested client and its limits.

Full setup, manual connection examples and troubleshooting: QUICKSTART.md.

What your AI can do

The hosted MCP service was checked on September 17, 2026: capability v64 · server 0.39.0 · 23 tools. Use the live capabilities tool for the current list and versions. Capability reference.

Bring over selected history

After connection, you can choose supported local history, a destination space and an extraction provider, then inspect the exact preview before accepting an initial import. Ongoing capture requires a separate setup and grant for the sources you choose. Connecting MCP alone does not read history or enable capture.

Source support and extraction-provider support are separate. The current qualified extraction path uses a version-specific Claude CLI profile; OpenAI and Cursor extraction providers are unavailable, and native Cursor history qualification is limited. Check the current history guide before choosing a source.

Selected source content is processed by the chosen extraction provider. ClariLayer stores the extracted context rather than a raw transcript archive. Provider processing may use your allowance or incur provider usage charges.

Optional semantic search has separate organization consent and provider disclosure in Settings. A new connection or history-import choice does not grant that consent. Lexical recall remains available without it.

Analytics when your work needs it

ClariLayer for Analytics retains its specialist tools:

• get_analysis_context recalls definitions, schema notes, saved queries and related context. Recall previews can be truncated; use get_context_entry for full bodies, stored SQL and CRM contracts.

• bootstrap imports supplied SQL, dbt models, notes, data dictionaries and semantic models. It does not read a repository or warehouse by itself.

• reconcile compares a saved definition with supported warehouse or bounded, row-free HubSpot evidence supplied by your agent. Salesforce contracts can be stored and recalled; Salesforce reconcile is disabled.

General work and engineering context are not independently reconciled. Live trust statuses remain asserted / caveat; verified is not live.

A seeded Analytics demo. Your agent supplies the warehouse evidence; a mismatch becomes a caveat for later sessions. This illustrates the Analytics specialist path.

Check dbt documentation locally

cd your-dbt-project
dbt docs generate
npx clarilayer dbt-check

The CLI compares local manifest.json and catalog.json files for phantom columns, missing catalog models, type-family mismatches and empty descriptions. No account or warehouse connection is needed for the local check. With an explicit --save, bounded findings are staged as proposals in your Context Inbox for review. --save --dry-run previews the payload without a network request.

See the CLI reference and Analytics recipes. The August 2026 dbt drift survey retains its method, evidence and limitations for the published 0.2.1 check.

Control and privacy

• Saved context retains its source and applicability. Separate spaces do not merge automatically.

• History import, ongoing capture and semantic processing each have their own setup or consent boundary.

• In the personal MCP reconcile path, ClariLayer holds no warehouse or CRM credentials, executes no SQL and calls no HubSpot API. Your agent supplies the evidence. Warehouse evidence may include optional preview rows; CRM evidence is row-free.

• Suggestions can be staged with propose / propose_batch for review. Pending proposals are not live recall context. Ad hoc conversation harvest requires an explicit request and sends distilled candidates, not the transcript.

Security · Privacy · Pricing

What is open source?

This repository's docs, examples, recipes and setup/dbt CLI are MIT licensed. The hosted ClariLayer service is proprietary. This repository is the public starting point for connecting to the service.

The Governed Context Edge is available through a hand-run private team pilot. Request early access.

For contributions and setup help, see CONTRIBUTING.md.