Single OriginDeep context graph for AI agents
Agents that knowyour warehouse
Faster answers. Right answers.
Single Origin turns your query history into a deep context graph: what runs, how it is built, what it costs, and who depends on what. Your agents query it over MCP, in the tools you already use.
Snowflake
Databricks
BigQuery
Redshift
Spark> Why did nightly_revenue_job cost 3× more this week?
⏺ singleorigin.job_run_comparison (nightly_revenue_job · last week vs this week)
⎿ cost 3.1× · bytes scanned 40× · same row count
⏺ singleorigin.query_profile_detail (costliest query of the latest run)
⎿ full scan of analytics.orders_fact · 0 partitions pruned
⏺ singleorigin.table_detail (analytics.orders_fact)
⎿ order_date is VARCHAR · clustered on order_date
The date filter no longer prunes: order_date is now a string, so the filter compares text and every run scans the whole table. Cast the filter to a date and scans drop back to about 2% of the table.
✓ 3 tool calls · 14 days of runs · query profiles
Trusted by data teams at
Never reads your rows
Read-only access to metadata and query history only.
SOC 2 Type II
Certified since October 2024.
Penetration tested
Third-party test passed in December 2025.
Encrypted everywhere
At rest with AWS KMS keys, TLS 1.2+ in transit.
Your tokens, your scope
Agents act with each user’s own access token.
Built for enterprise data teams. Read about security →
Agents can write queries. They don't know your warehouse.
An agent can read your repo, but the facts that matter for data work live in production: which queries run, how often, what they cost, and what depends on what.
No production memory
Without execution history, an agent optimizes the query in front of it and misses the hundred near-copies running every day.
Raw logs don't fit
A large company logs terabytes of query and job data a day. Feeding it to an LLM is slow, noisy and expensive.
Guessing isn't safe
Dropping a column or rewriting a pipeline needs exact answers about who reads it, not a plausible one.
Same question, two agents
“Why did nightly_revenue_job cost 3× more this week?” Both agents can write the query. Only one can see what actually ran.
Without Single Origin
Reads the repo and guesses.
- 1.grep -r "orders_fact" across the repo
- 2.reads 6 SQL files, 2 of them dead code
- 3.guesses which one runs at night
- “Try adding more compute to the warehouse, or a clustering key.”
With Single Origin
Looks up production history and cites it.
- 1.job_run_comparison: scans up 40× since Tuesday
- 2.query_profile_detail: 0 partitions pruned
- 3.table_detail: order_date is now VARCHAR
- “Cast the date filter back to a date. Scans return to about 2% of the table.”
Give your agents enterprise memory
Single Origin continuously builds a context graph from your production query history and serves it to agents as MCP tools, so they answer with evidence instead of guesses. Here is what they can ask.
Hover or tap a node. Example graph; names are made up.
Table
analytics.orders_fact
- •1.2B rows · clustered on order_date
- •Written by 1 job · read by 38 jobs
- •Most reads scan 2% of the table
Your agent can ask
“What breaks if I change this table?”
Query profiles
“Why is this query slow, and which step costs the most?”
Execution history and compacted plans for every query pattern.
See the toolsSimilar queries
“Where else does this logic run?”
Queries grouped by what they actually do, across the whole workload.
See the toolsTables and jobs
“What builds this table, and why did its job cost more this week?”
Producers, build chains, scan cost and run history from real queries.
See the toolsRecommendations
“What should we fix first?”
Ranked fixes, each with the evidence behind it.
See the toolsQuestions agents ask Single Origin
- “Why did this job cost more this week?”
- “Which step of this query costs the most?”
- “Where else does this logic run?”
- “What builds this table?”
- “Is it safe to drop this column?”
- “What should we fix first?”
- “Did last month’s fix actually save anything?”
- “How much of this table do readers scan?”
- “Which tables are built the same way?”
- “Which jobs are far above their usual day?”
Context wherever data work happens
No new UI to learn. The same context reaches your agents, your pull requests and your backlog.
In your coding agent
Claude Code, Cursor, Claude Desktop or any MCP client. Go from question to merged PR with evidence from production.
Connect your agentIn code review
A GitHub Action reviews query changes in pull requests against how your warehouse actually runs, using your team's rules, and comments inline.
How code review worksIn your backlog
Ranked optimization recommendations, refreshed daily, that an agent can pick up and turn into a change.
Browse recommendationsWhy not build it yourself?
A RAG pipeline over query logs is easy to start and hard to trust. Good agent context has to be distilled, current and correct.
Distilled, not dumped
Native parsers for each warehouse dialect and clustering reduce hundreds of thousands of daily queries to patterns, and compact each plan to the parts that matter.
Incremental
Only new query patterns are analyzed each day, so context stays current without re-reading everything.
Checked against evidence
Usage is mapped at table and column level, and every recommendation is checked against query profiles before an agent sees it.
Context at Coinbase scale
Tens of thousands of tables, hundreds of thousands of Snowflake queries a day, and queries over 1,000 lines long.
of queries parsed, up from 30–40%
of queries on the top 100 tables can be optimized
parsing and optimization accuracy on covered queries
“Single Origin was the only third-party option … offering all of them at scale and meeting our data security requirements.”
Built by Data Infrastructure Veterans
Our team built and scaled data infrastructure at the companies that defined modern data engineering.
Give your agents production context
Connect Single Origin to your MCP client and ask your first question about your own workload.