Understand your codebase.
From planning to review.
Meet our agent.
It does not grep your codebase, it queries it from a compiled model.
ExportCSV in invoices/export.go casts Seats to an int, so a 2.5-seat upgrade exports as 2 and the export stops matching the invoice. RenderLineItems and MonthlySeats keep the fraction and are fine.
Two tests call ExportCSV, and both use whole seat counts, so neither would catch the 2.5-seat export. TestExportCSV_Upgrade is the place to add that case.
One agent, wherever you ask.
People and coding agents talk to the same agent, which queries the same compiled model, so nobody works from a different picture of the code.
- In the app Beside every review, wiki page and check.
- In Slack In a direct message or a thread, where the team already talks.
- In pull requests Ask in a comment and the answer lands on the PR.
- In your coding agent Claude Code, Codex and Cursor can ask our agent over MCP instead of searching files.
- In your terminal Through the abstr command line.
In the app and Slack, the agent drafts checks and proposes concept edits as cards. Nothing changes until someone approves it.
The same coding agent, with and without Abstraction.
Same model and the same commit of our monorepo on both sides, with every answer graded against a fixed key.
Questions about structure
What eventually calls this function? Which functions call the most others, or run the longest?
- Several times faster
- A fraction of the cost
- Correct in every run
Questions about how something works
How does a request get authenticated? What marks a record as dirty?
- About as fast
- Cheaper
- Right more often
The widest gap: find every function that eventually calls one worker entry point.
Reason once, reuse everywhere.
Concepts are the understanding above the code. An agent works it out once and writes it down. Every run after that reads the result, and the result never drifts.
- 01 Read 38 files
- 02 Traced 112 calls
- 03 Followed 9 data flows
- 04 Wrote it down
Prorates seat changes to the day for monthly subscriptions. Annual plans never prorate.
Cycle length comes from the subscription, not the plan, since PR #4821.
- The agent, asked in Slack 2 × 1,000
- The review of PR #4821 3 × 1,000
- Claude Code, planning through MCP 5 × 1,000
A concept can be a subsystem, a process, a glossary term, or the context around a piece of code that never fit in a comment. It is compiled and re-verified from the code on every change, so it never describes last quarter’s system. The wiki stays for people.
Diagrams compiled from the code.
Architecture maps and control flow diagrams, compiled from the code itself, so they cannot drift. The readability of a low-code interface, for the code you already have.
How it works
A semantic model of your code, powering every PR review.
An abstract syntax tree compiled from each supported language in your stack. Callgraphs, control flow, and data flow derived from it. Semantic diffs apply incrementally as code changes, and every PR is reviewed against the whole model.
Compile
Every language, one semantic representation.
We compile every line of code in every service into the same abstract syntax tree. Algorithmic, deterministic, no LLM in the loop.
This is the foundation everything else is built on. We call it the Universal Syntax Compiler.
Why Abstraction?
Because you can't sign off on what you don't understand.