Skip to content

Coding agent · made in the Caribbean

Take it easy. The code gets done.

A coding agent in your terminal. It reads, searches, edits, and runs commands in your project, with your permission.

npm install -g quillami-code
Get started

Node 22 or newer · MIT license · the CLI speaks Spanish, the model answers in your language

Sample session: Quillami fixes a failing test.

Vamo' a darle, que el código no se escribe solo.
› fix the failing date test
· jev · estándar → MiMo V2.6 Pro · $0.435 / $0.87 por M
· bash npm test
✔ ¿Qué hago? Sí, y no preguntes más por bash en esta sesión
  exit 1 · 3.1 s · 42 líneas
· edit src/date.ts
✔ ¿Qué hago? Sí, solo esta vez
· bash npm test
  exit 0 · 2.8 s · 38 líneas
The test expected the date in UTC and the function used local time. It formats in UTC now.
  tokens: 21.4k in · 640 out · sesión 22k · $0.01 esta sesión
›

You give it the task. It finishes it.

Quillami picks the tool, asks before it changes anything, reads the result, and keeps going until it is done. Ctrl+C cancels the turn, not the session.

  1. task

    Say what you want, in your own words. One-shot works too: quillami "explain this file".

  2. tools

    read, grep, glob, and ls to look. write, edit, bash, and web_fetch to act.

  3. permission

    Anything that changes something asks first, with the diff in front of you. Yes, no, or yes for the rest of the session.

  4. repeat

    It reads the output, fixes, and tries again. Up to 20 rounds per turn.

Auto mode

The cheapest model that clears it.

For every message, Jev (by TypeSafe) rates how hard the task is, and Quillami uses the cheapest model that handles it well. A quick question does not pay Opus prices.

Models auto uses through OpenRouter. US dollars per million tokens, input / output.
DifficultyModelPriceFor
TrivialDeepSeek V4.1 Flash$0.03 / $0.50A short question, one step
StandardMiMo V2.6 Pro$0.435 / $0.87Everyday coding in the repo
HardClaude Sonnet 5.5$2 / $10Design, many files, deep debugging
Very hardClaude Opus 5.5$4 / $20System architecture, subtle bugs across modules

With a free Artificial Analysis key, auto picks from the whole OpenRouter catalog by benchmark.

Without a Jev key it uses MiMo V2.6 Pro for every message. With an Anthropic key it uses Haiku, Sonnet, and Opus directly.

/usage shows what you spent and what you saved compared with always using Claude Sonnet.


› /usage
  turno    14.2k in (81% de caché) / 520 out  $0.0041
  sesión   186k in (74% de caché) / 6.3k out  $0.12
  auto     DeepSeek V4.1 Flash ×11 · MiMo V2.6 Pro ×5
           $0.12 vs $0.84 con Claude Sonnet siempre (ahorro 86%)
/usage output from a sample session.

Permissions

See the diff before you approve.

write and edit show what goes out and what comes in before touching the file. Pick with the arrows and Enter, or press s, n, or a. Esc means no.

  • Jev approves only harmless commands and stops high-risk ones.
  • “Don’t ask again” covers only that tool: approving bash does not approve write.
  • In plan mode it only reads and ends with a plan. Nothing changes.
· edit src/date.ts · cambia un bloque

@@ -12,3 +12,3 @@
 export function format(date: Date) {
-  return date.toLocaleDateString("en-US");
+  return date.toLocaleDateString("en-US", { timeZone: "UTC" });
 }


? ¿Qué hago?
❯ ◯ Sí, solo esta vez
  ◯ No, no lo toques
  ◯ Sí, y no preguntes más por edit en esta sesión
  ↑↓ moverse · espacio marcar · enter elegir · esc cancelar
How Quillami asks before editing a file: yes once, no, or yes for this session.

What it brings.

What you need to work on a real repo, with nothing to configure beyond your keys.

  • /skills

    Skills

    Uses the skills you already have in .claude/skills or .agents/skills. It loads the one that fits, or you force one with /name.

  • /model

    OpenRouter

    Hundreds of models with one key. /model opens the list and you search by typing.

  • /mcp

    MCP

    Connect stdio or HTTP servers in ~/.quillami/mcp.json.

  • quillami -c

    Sessions

    Pick up the last conversation in this repo. /sessions lists the rest.

  • /mode

    Modes

    agent asks, plan only reads and proposes, yolo approves on its own. The denylist and Jev still apply.

  • QUILLAMI.md

    Memory

    QUILLAMI.md or AGENTS.md for the repo. ~/.quillami for what it learns about you.

  • web_fetch

    Safe web

    Reads public pages. Blocks localhost, private networks, and redirects to them.

  • Ctrl+C

    Cancel freely

    Stops the turn, not the session. /undo reverts the last turn’s changes.

Install it.

npm install -g quillami-code

  1. Install with npm. You need Node 22 or newer.

  2. Run quillami in your project. The first time, check your keys with the space bar: OpenRouter and Jev are recommended.

  3. Type the task. Quillami asks before it touches anything.

Get your keys

You need at least one model key: OpenRouter or Anthropic. Quillami asks for them the first time and saves them in ~/.quillami/.env.

  • OpenRouter

    Recommended

    Hundreds of models with a single key. This one is enough to start.

    OPENROUTER_API_KEY
  • Jev (TypeSafe)

    Recommended

    Picks the model for each message and approves only harmless commands.

    TYPESAFE_API_KEY
  • Anthropic

    Optional

    Claude directly, without going through OpenRouter.

    ANTHROPIC_API_KEY
  • Artificial Analysis

    Optional · free

    With OpenRouter and Jev, auto mode picks by benchmarks.

    ARTIFICIAL_ANALYSIS_API_KEY
Read the full guide