Schemas change, definitions move, and nobody edits the context repo. Cassis catches the drift, drafts the fix, and proves it against evals. So your agents are always served current, approved context.
First answer in about two minutes. Read the quickstart
01/06 · Listing maintenance issues
02/06 · Drafting the metric proposal
description: > Accounts whose first successful production deployment occurred in the requested period. Internal accounts excluded. display_name: Activated accounts domain_path: product/activation expression: COUNT(DISTINCT "ACCOUNT_ID") filters: '"IS_INTERNAL" = FALSE' name: activated_accounts synonyms: - activated accounts - activated customers table_name: FIRST_DEPLOYMENT table_schema: PRODUCT unit: accounts
03/06 · Validating the context files
04/06 · Testing against the eval suite
05/06 · Pushing the review branch
06/06 · Opening the GitHub pull request
description: > Accounts whose first successful production deployment occurred in the requested period. Internal accounts excluded. display_name: Activated accounts domain_path: product/activation expression: COUNT(DISTINCT "ACCOUNT_ID") filters: '"IS_INTERNAL" = FALSE' name: activated_accounts synonyms: - activated accounts - activated customers table_name: FIRST_DEPLOYMENT table_schema: PRODUCT unit: accounts
How many accounts activated last month?
accounts.status = 'active' 2,431 accounts activated last month.
first_deployment.status = 'succeeded' 1,847 accounts activated last month.
MEDIAN(deployments.duration) 3 context references affected
As the context runtime for your analytics agents, Cassis records a gap each time an agent has to guess: no governed definition, two plausible signals, a clarifying question.
It also diffs every schema and dbt change against the context that references it. Your dbt definitions are imported, not redefined.
Cassis structures your context as a Markdown and YAML file tree following Google's Open Knowledge Format.
Use the CLI and MCP from your preferred development environment to inspect detected issues, review and test proposed fixes, and publish approved context through your existing Git workflow.
acme-data wants to merge 1 commit into main from cassis/activated-accounts
Evidence: 6 observed conversations · Proposal only
The repository above is a specimen. Browse a real context repo on GitHub, CI gates included.
Cassis uses your schema, dbt project, documentation, and questions. Your warehouse or existing agent executes the generated SQL. Row data stays in your environment. Evals score by SQL equivalence review.
Cassis executes generated SQL through a dedicated read-only role. Access is limited to the datasets your agents use, with no warehouse writes. Evals execute the generated and expected SQL and compare result rows.
Hosted on AWS in Paris, in the eu-west-3 region.
A Cassis-managed private deployment in your own cloud environment.
Cassis manages Anthropic inference through Amazon Bedrock.
Use your Amazon Bedrock or Google Vertex AI account.
Connecting your own warehouse starts with a conversation. One domain's schema is enough, as a read-only connection string or a DDL dump. Talk to us.
Connect Cassis to your Git repository and install the Cassis CLI. Then give it your schema: a scoped read-only warehouse connection, or a DDL dump if you would rather not connect one yet.
Run the bootstrap from your coding agent. Cassis writes the first context tree to your repository from your schema, dbt project, BI exports, query history, and docs.
Connect Claude Code or any MCP-compatible agent. Cassis can answer questions, generate SQL, or return only the context the calling agent needs.
List detected issues from your development environment, apply and test proposed fixes, then merge the approved pull request to publish a new context version.
Run the full loop now on Stallora, the sample marketplace every new account starts with, or bring one domain from your own stack.
Already live in prod at leading tech companies, from a few hundred to a few thousand employees.
Your stack already holds much of the context an analytics agent needs. Here's how to recover what's there without inventing what isn't.
GitLab, Mattermost, Cal-ITP: 13 public dbt projects and 5,284 models. Here is the context their metadata still cannot establish for an analytics agent.
Rules, metrics, or worked examples? We measured which context representations shorten an LLM's path to a correct answer, and what the wrong turns taught us about good context.
