Write it. Run it. Ship it.
A little natural language. A little Python.
A program that's yours.
uv add kedi
A PROGRAMMING LANGUAGEtmdb.kedi · Tool source A small install.
A different way to program.
41.0082° N / 28.9784° Ekedi
Program with natural language.
Typed by design.
Compose model calls, decisions, and tools as typed programs. Natural language is part of the code, not a string on the side.
movie_night.kedi
# Load the movie lookup tool.> import: tmdb# Let the model use it.> approval: allow> use: movie_details # Find the film by name and capture its details.>> Look up Spirited Away using movie_details. Its director is [director: str], release year [year: int], and runtime [minutes: int] minutes. # Use the result as a program value.= <director>BOUND VALUESIllustrative output
- director
- str
- Hayao Miyazaki
- year
- int
- 2001
- minutes
- int
- 125
uv add kediA different way to program.
01 / THE LANGUAGELESS GLUE. MORE MEANING.
From a sentence
to a typed value.
The context, the question, and the shape of the answer.
All in the same line of your program.
>> The next action in is .INPUT /
note: strQA must reproduce the timeout before Mira can fix checkout.
next_action.kedi
~Action(owner: str, task: str, blocked: bool) [note] = QA must reproduce the timeout before Mira can fix checkout.>> The next action in <note> is [action: Action]. = `action`SENTENCE IN. VALUE OUT.
Explore templates Not a string you have to untangle.
Use the result in the next statement, pass it to a procedure, or hand it to Python.
02 / NATURAL CONTROL FLOWWORDS CAN BE CONDITIONS, TOO.
If it needs attention.
While it needs work.
Let a model evaluate a claim.
Let your program decide what happens next.
release_gate.kedi
[change] = Rename user_id to account_id in API responses.[track] = standard > if: <change> requires existing integrations to change [track] := migration = <track>BOUND VALUESIllustrative output
- track
- str
- migration
ONE CLAIM. TWO PATHS.
Does this API change
need a migration?
Truemigration
Falsestandard
Natural-language claims and deterministic Python conditions. Same control flow.
grounded_copy.kedi
[facts] = CSV export ships Monday. PDF is not available.[draft] = Export any report as CSV or PDF today. > loop: <draft> makes promises not supported by <facts> >> A revision of <draft> supported only by <facts> would read [revision]. [draft] := <revision> = <draft>BOUND VALUESIllustrative output
- draft
- str
- CSV exports arrive Monday.
CHECK. REVISE. CHECK AGAIN.
- 01Check the promises
Does the copy claim more than the facts support?
- 02Revise against the facts
Remove unsupported formats and launch dates.
- 03Evaluate the new draft
Stop when no unsupported promises are found.
The condition is evaluated again before every iteration.
issue_inbox.kedi
[reports: list[str]] = `["CSV loses the last row.", "Please add a dark theme."]`[bugs: list[str]] = `[]` > loop [report]: `reports`: >> It is [is_bug: bool] that <report> describes a defect. The issue title is [title: str].> map: > if: `is_bug`: `bugs.append(title)` = `bugs`BOUND VALUESIllustrative output
- bugs
- list[str]
- ["CSV export drops the final row"]
ONE CONTINUATION PER ITEM.
- 01Start with a collection
Each incoming report gets its own iteration scope.
- 02Continue when its answer is ready
The map receives that report's classification and title.
- 03Keep what qualifies
Collect bug titles; leave feature requests out.
Independent continuations can overlap. Kedi waits for the map stage before moving on; completion order may differ.
03 / A LANGUAGE, NOT AN ISLANDKEEP YOUR PYTHON CLOSE.
Keep writing your program.
Templates are the beginning. Compose them with the tools you already use.
quote.kedi
[quote] = 12 seats at 19 dollars per month, billed yearly. >> In <quote>, there are [seats: int] seats at [monthly_usd: float] dollars per seat per month. [annual_usd: float] = `seats * monthly_usd * 12`= `annual_usd`BOUND VALUESIllustrative output
- seats
- int
- 12
- monthly_usd
- float
- 19.0
- annual_usd
- float
- 2736.0
A backtick away
from Python.
A model reads the pricing terms. Python multiplies the typed numbers to get the annual total.
on_call.kedi
~Brief(problem: str, next_step: str) @handoff(report: str) -> Brief: >> The on-call handoff for <report> is [brief: Brief]. = `brief` [report] = Checkout times out only after a coupon is applied.[brief: Brief] = <handoff(<report>)> = `brief`BOUND VALUESIllustrative output
- brief.problem
- str
- Checkout times out with a coupon
- brief.next_step
- str
- Reproduce checkout with and without a coupon
A workflow
you can call.
Turn a report into a reusable, typed on-call handoff: a problem and a concrete next step.
stock_check.kedi
```import jsonfrom pathlib import Path``` @stock(item: str) -> int: = ``` path = Path("inventory.json") return json.loads(path.read_text())[item] ``` > approval: allow> use: stock>> The stock tool reports [remaining: int] units of "field-notebook". It is [can_fulfill: bool] that this can fill an order of 4. = `can_fulfill`BOUND VALUESIllustrative output
- remaining
- int
- 3
- can_fulfill
- bool
- False
Tools, in context.
The model checks actual stock before answering. A Kedi procedure reads the inventory through Python.
Example inventoryTyped boundariesLexical scopesPython interopPluggable agent adapters
04 / PUT THE PIECES TO WORKONE PROGRAM. A SMALL TEAM.
From building blocks to a working crew.
A researcher finds the breaking change. An editor explains the migration. Profiles, tools, subagents, and typed answers, in the same program.
release_team.kedi
> import: filesystem> model: google/gemini-3-flash-preview > profile: researcher: > instructions: Read CHANGELOG.md. Find required v2.0 migrations. > use: read_text_file > output: list[str] > profile: editor: > subagent: researcher > use: editor>> The v2.0 migration checklist reported by the researcher is [steps: list[str]]. = `steps`BOUND VALUESIllustrative output
- steps
- list[str]
- ["Read account_id instead of user_id in API responses"]
05 / KEDI + JEVJEV BY TYPESAFE
Typed judgments.
Your next move.
Pick a queue. Assess cancellation intent. Let your program decide the priority.
ticket_triage.kedi
> import: decisions> model: typesafe/jev-latest [ticket] = I was charged twice. Reverse it or I cancel today. >> <ticket> belongs to [queue: Literal["billing", "technical"]]. The probability the customer intends to cancel is [churn: Probability]. = `{"queue": queue, "priority": churn >= 0.80}`BOUND VALUESIllustrative output
- queue
- Literal
- billing
- churn
- Probability
- 0.93
- Choice
- A
Literallimits the answer to your defined options. - Probability
- A value from 0 to 1. Keep the assessment; set your own decision rule.
06 / INTERACTIVE KEDIA LITTLE ROOM TO EXPERIMENT.
Your program, one step at a time.
One cell leads to another.
Try a template. Inspect its values. Keep going in the next cell, with your program's state right where you left it.
kedi notebookOpens your browser at localhost:8788/notebook/. Served locally, on your machine.
First-time setup
From a Kedi source checkout with its submodules initialized, install the notebook dependencies and launch it with:
uv run --extra notebook kedi notebookAfter installation, run kedi notebook with that environment active.
NotebookCity notes[1] > model: openai/gpt-6-luna >> In Türkiye, [city: str] spans the [strait: str]. [2] > show: <city> spans the <strait>. Istanbul spans the Bosphorus Strait.
0:09 / 0:09
Your Python, your choice.
Start with browser Python, or select a host interpreter with a reusable notebook virtual environment.
More than a scratchpad.
Mix Kedi, Markdown, and terminal cells. Save a .kedinb notebook, with optional execution progress and no Secret Manager values.
Stay in the terminal.
Write a template, inspect a capture, and build on it. Variables, procedures, and model settings stay with your session.
kedi --idleSyntax highlighting is on by default. Use --no-color for plain text.
$ kedi --idle +++ > model: openai/gpt-6-luna +++ >> [city] is the city of cats. +++ > show: `city.upper()` ISTANBUL +++ > show: Cats love <city>. Cats love Istanbul. +++
0:11 / 0:11
Start small. Keep going.
Use :multiline for a longer fragment and :show to inspect a value.
Take it with you.
Save your session's source to a .kedi file with :dump.
COMING SOON
Kedi Harness
An agent harness.
With a mind of its own.
Meet the cat taking the keyboard. An agent harness for your projects, with Kedi as its orchestration language. It can write and run Kedi programs to coordinate its own tools, subagents, and next steps.
Follow the work on GitHub
07 / THE TEAM
The people behind Kedi.
Mert Sırakaya
Co-creatorApplied AI Engineer
Doğukan Yiğit Polat
Co-creatorPhD Student
08 / YOUR FIRST .KEDI FILE
KARAKÖY
KADIKÖY
