Software
intelligence
that stays.

An AI-native engineering environment where every fix, every review, every rule becomes intelligence your team owns — not the model's.

v0.1.5 · macOS 13+ (Apple Silicon, signed & notarized) · Windows 10+ (x64)

The premise

Every model forgets by morning. Every IDE rebuilds the same understanding from scratch. Tavsin doesn't — it remembers the system, the standards, and the specific reason last quarter's fix broke staging.

Understand

It reads your system, not just your file.

On a single scan, Tavsin builds a live intelligence graph of your repository — files, functions, routes, schema, tests, dependencies, and the git history that produced them. It knows what's fragile, what depends on what, and which tests actually cover a change.

Read about the graph
~/repo $ tavsin scan
 412 files · 1,847 symbols · 96 routes · 218 tests
 dependency edges resolved (NodeNext)
 fragility scored from 4,108 commits

~/repo $ tavsin fragile --top 3
src/auth/session.ts      0.91  churn × fan-in
src/billing/invoice.ts   0.84  no tests · 7 dependents
src/queue/dispatcher.ts  0.79  3 rewrites this quarter

~/repo $ tavsin impact src/auth/session.ts
blast radius: 23 files · 12 tests cover
 middleware/requireAuth.ts
 api/login.ts, api/refresh.ts
 ... 20 more
goal:          fix the flaky login regression
success:       login works under concurrent sessions
                and the regression is covered by a test

verification:
  - tests:    npm test
  - contract: no new TODOs · no skipped tests
  - review:   summary references the failing case

evidence_required:
  - test output
  - the file or files changed
  - a one-line memory: "why this fix is safe"

status:        satisfied  ·  3 of 3 checks passed

Contract

It writes a contract before it writes the code.

No serious outcome ships without success criteria. Every workflow begins with an Outcome Contract — what done means, how it's verified, the tests, the risks, the evidence. The enforcer reads the trace and decides whether the work actually shipped, against the contract you wrote.

Read about outcomes

Compound

It keeps what it learns.

Every workflow produces a trace. Every approved rule, fix, and convention becomes a memory the next agent surfaces automatically. The intelligence compounds in a local store you can inspect, correct, export, and own — not vanish into a vendor's training data.

Read about memory
memory · approved added 14 days ago · surfaced in 7 runs

auth.ts is fragile — keep a regression test

Two prior incidents traced back to session.ts being edited without running the concurrent-login case. Don't merge changes to this file unless auth.regression.test.ts is green and references the originating ticket.

The metric

We don't count tokens. We count what compounds.

Lines of code measure activity. Token spend measures appetite. Neither measures whether your team got better at shipping. Tavsin's core metric is simple: Verified Outcomes — work the enforcer proved against tests and evidence — plus Reusable Learning Created, the rules and fixes your team kept.

Principles

Built sovereign by default.

Local-first.

All intelligence lives in a local SQLite store on your machine. Inspectable, exportable, deletable. Deletes are soft so your audit stays complete.

Model-agnostic.

The model is not the moat. Switch between Anthropic, OpenAI, Gemini or your own local endpoint without touching a single trace, memory, or skill.

Open by design.

Every meaningful run leaves an append-only trace — training signal you can read, not log noise you can't. The repository is open source.

Get started

Three commands. A real project.

$ brew install tavsin # or download from releases $ tavsin init && tavsin scan $ tavsin run "audit the riskiest files"

Free to run. Local by default. Read the source.