A developer
In a million-line application with no docs, finds the routes one by one, reads what each takes and returns, and writes it down. It takes weeks, and goes stale as the code changes.
Sezgi is Turkish for intuition: knowing without thinking.
A sense of self for software.
Sezgi reads the application it is installed on, whatever language it is written in, once, and turns it into one API. The AI stops scanning code and asks Sezgi; every answer comes with its source and a stamp saying whether it is certain or a guess. Writes happen only as far as you allow, and always through the application's own rules.
Open source (Apache-2.0) · fully local, offline · no language model inside · 0.1, experimental
Sezgi is not an application, it is a layer. One can of paint covers a whole building, and a building holds thousands of cans.
The problem
To connect an application to AI today, an API is written, documented and wrapped in an MCP server, separately for every application. Most companies never did it.
An AI
Is a stranger to the application. Opens files, guesses, and starts over for every question. Hours and millions of tokens go by; the result is often incomplete or wrong.
People do not work like this. Asked their name, they say it; they do not wonder who they are. They learned it once and it became part of them. Sezgi gives software that sense of self.
Who it is for
Developers
Walking undocumented code for weeks.
The map, OpenAPI document and Benlik Card come out of the code by themselves; when code changes, only the difference is re-read.
AI assistants and agents
Re-scanning code every session, making wrong calls.
They ask and get short answers with sources. A wrong write is stopped at the door, in the preview.
Enterprises
Undocumented legacy systems, months-long integrations, an auditor asking "what did the agent touch?"
The map is there on day one. Writes go with the real user's identity, through the application's checks, and every action is logged.
SaaS companies
Customers asking "can I connect my AI?"; integrations breaking every release.
A ready MCP server and up-to-date OpenAPI. Changes that would break customers are caught before release.
AI companies
Paying for their agents' wrong actions.
The agent reads the card, not the code: accuracy goes up, LLM tokens and GPU waste go down.
Agencies and integrators
Learning each new client's system from scratch.
The map is there on day one; the coverage report goes into the proposal, the gaps found go to the client.
Application owners
Not seeing what they open to AI.
The Ayna screen shows what is known and what is not; they choose the permissions and keep the last word.
Security and audit
Undocumented doors and unchecked endpoints going unnoticed.
In ten minutes, without granting anyone access, they see writes that need no login and endpoints with no access check found.
How it works
sezgi initreads the code and maps operations, inputs, rules, access, tables, outside effects and reversibility. It never writes into the application's code.sezgi connectadds Sezgi to your AI tool (MCP), asking tool by tool;sezgi removeundoes it.- Ask: "If I delete a user, which tables are affected, and can it be undone?" The answer comes with its source and a trust stamp.
Optionally, live mode: the assistant uses the application, on a test copy first. Every request passes permission, preview and effect checks; a delete, an outside effect or anything uncertain needs human approval.
The parts of the self, and today
These twenty points are the project's contract. Next to each is its honest state in this release.
- Knows itselfPartly
- Reaches directly, never scansYes
- Covers everythingPartly
- Learns repetition oncePartly
- Does not think, knows; never makes things upYes
- Knows and enforces the rulesPartly
- Is one API (MCP, HTTP, OpenAPI)Yes
- Use the app like a databasePartly
- Masters the interfaceWeak
- Issues tokens, you set the permissionsYes
- Single counterpart, a bridgeYes
- Does not discriminateYes
- Works in every languagePartly
- Almost no memoryPartly
- Tries to be unbreakablePartly
- Removes wastePartly
- Is a layer, leaves no traceYes
- Goes beyond the app, if askedStarted
- Does no harmYes
- Publishes a Benlik CardYes
Promises that do not change
Never makes things up
Every fact comes with file and line, stamped certain · observed · inferred · unknown.
Grants nothing, only narrows
Effective permission = the user's permission in the app ∩ the token's permission. No human, no write.
Asks when unsure
Where unsure, the answer is approval. On real data every write needs the human's password.
Local
Code and data never go to an outside service; no telemetry.
Keeps no customer data
Not written by default; everything kept is declared in the data ledger.
Evidence first
Every accuracy and speed claim is backed by an open measurement.
Benlik Card
Every application introduces itself with a signed, open card: what it can do, which rules it follows, on whose behalf writes happen, and what is not certain. The first API description that can say "I am not sure". The specification is open: benlik/SPEC.md.
The goal
A connected assistant should answer 17 founding questions about an application it has never seen, within its first minute, correctly, with sources and honestly: what it is, why it exists, who uses it, who may do what, which tables and systems are affected, whether it can be undone, whether a human must approve… Today Sezgi answers 2 reliably and 14 partly on never-seen apps. The target is above 90% accuracy and zero wrong claims stamped "certain".
Roadmap
| Code reading, map, Benlik Card, MCP, Ayna | Released |
| Writes on a test copy, approval queue, identity delegation, data ledger | Mostly there |
| Switchboard joining many apps in one connection | Yes |
| Interface: screens and the operation each button starts | Started |
| Internal code: calling functions without an API, isolated | Not yet |
| 1.0: independent security audit, writes on real data | In progress |
| Beyond the application: command line, operating system, programs without source | After 1.0 |
Try it
cd path/to/your/app
sezgi init
sezgi connectReady binaries for Windows, macOS (Apple Silicon and Intel) and Linux (x64, ARM). Latest release · Install and commands