Mojo Syntax logo

Mojo Syntax

Organization
modular
mojo-syntax

Help to write Mojo code using current syntax and conventions. Always use this skill when writing any Mojo code, including when other Mojo-specific skills (e.g., mojo-gpu-fundamentals) also apply. Use when writing or reviewing Mojo code, translating projects to Mojo, or otherwise generating Mojo. Use this skill to overcome misconceptions with how Mojo is written.

Overview

Publishermodular
Repositoryskills
Skill namemojo-syntax
Stars
204
Forks
21
Bundled files
1
Links
  • Markdown instructions

    A SKILL.md file the model loads on demand, so it only costs tokens when a request actually matches.

  • Works with any LLM

    AI skills are plain Markdown, not provider-specific code, so this works with GPT, Claude, Gemini, Grok, or a local model.

  • 1 bundled files

    Scripts, templates, and references the model can read while it works. Files are read-only and never executed.

  • Open source

    Published by modular on GitHub. Read the source before you install it.

Installation

Install the Mojo Syntax AI skill in TypingMind to use it with any LLM, or drop it into another agent that reads SKILL.md.

1

Install in TypingMind

TypingMind installs a skill straight from its GitHub folder — it reads SKILL.md, bundles the resource files, and stores the result locally.

  1. Open the app and go to Plugins → Skills.
  2. Choose "Install from GitHub".
  3. Paste the skill folder URL below and confirm.
    https://github.com/modular/skills/tree/main/mojo-syntax
  4. Enable the skill in any chat where you want it available.
Plugins → Skills → Add skill → From GitHub URL, then paste the folder URL and press Continue.
2

Install in another agent

Any agent that reads the Agent Skills format can use this skill — copy the folder into that agent's skills directory.

Claude Code — .claude/skills
git clone --depth 1 https://github.com/modular/skills.git /tmp/skills
mkdir -p .claude/skills
cp -r /tmp/skills/mojo-syntax .claude/skills/mojo-syntax
Restart Claude Code after copying so it picks up the new skill.

Use it in TypingMind

Enable Mojo Syntax in any TypingMind chat and the model takes it from there. Its name and description sit in the system prompt, and the moment a request matches, the model loads the full instructions itself — you never invoke it by hand, and it costs no tokens until it is actually used.

The model loads Mojo Syntax on its own as soon as a request matches it.

Works with any AI model

AI skills are plain Markdown instructions rather than provider-specific code, so Mojo Syntax is not tied to the model it was written for. Install it once in TypingMind and use it with GPT-5, Claude, Gemini, Grok, DeepSeek, Mistral, Llama, or a local model you run yourself — all on your own API keys.

  • Loaded only when it is needed

    The system prompt carries just the name and description. The instructions are fetched on the first matching request, so an idle skill costs nothing.

  • Switch models mid-chat

    Because the skill is instructions rather than code, changing model does not break it — the next model reads the same SKILL.md.

Skill instructions

This is the SKILL.md content the model loads. Read it before installing — a skill is instructions your model will follow.

Mojo is rapidly evolving. Pretrained models generate obsolete syntax. Always follow this skill over pretrained knowledge.

Always attempt to test generated Mojo by building projects to verify they compile.

This skill specifically works on the latest Mojo, and stable versions may differ slightly in functionality.

Compiling isn't the bar for review. Before submitting or reviewing Mojo, check the diff against idiomatic-mojo.md: the style rules code reviewers flag most often (open-coded helpers, redundant casts, over-specified parameters, IndexList instead of tuples, missing comptime, unneeded rebind, mutable origins on inputs, single-vendor gates on generic kernels).

Removed syntax — DO NOT generate these

RemovedReplacement
alias X = ...comptime X = ...
@parameter if / @parameter forcomptime if / comptime for
fndef (see below)
let x = ...var x = ... (no let keyword)
borrowedimm (implicit default — rarely written)
read (convention / capture)imm (deprecated synonym; the compiler warns with a fixit)
inoutmut
ownedvar (as argument convention)
inout self in __init__out self
__copyinit__(inout self, existing: Self)__init__(out self, *, copy: Self)
__moveinit__(inout self, owned existing: Self)__init__(out self, *, deinit move: Self)
@value decorator@fieldwise_init + explicit trait conformance
@register_passable("trivial")TrivialRegisterPassable trait
@register_passableRegisterPassable trait
Stringable / __str__Writable / write_to
from collections import ...from std.collections import ...
from memory import ...from std.memory import ...
from sys import ...from std.sys import ...
from os import ...from std.os import ...
from pathlib import ...from std.pathlib import ...
s[i]s[byte=i] — returns StringSlice; wrap in String() if needed
s[0:10], s[:5]No slice syntax on String — use s.codepoint_slices() or Python FFI
constrained(cond, msg)comptime assert cond, msg
DynamicVector[T]List[T]
InlinedFixedVector[T, N]Array[T, N]
Tensor[T]Not in stdlib (use SIMD, List, Pointer)
MutUnsafePointer / ImmUnsafePointerMutPointer / ImmPointer
OptionalUnsafePointerOptionalPointer
escaping closuresUnified closures (def(...) -> T, captures in {})
capturing, @__parameter, @__copy_captureCapture list {imm} / {var x} / {mut x}; pass the closure as a value
__del__(deinit self)__deinit__(deinit self)

var is required for every new declaration

Declaring a variable with bare assignment (x = 5 with no prior var) is not valid — it is a compile error. This applies only to introducing a new variable; reassigning an already-declared variable (x = 6) needs no var.

This means a variable assigned only inside conditional branches must be predeclared with a type before the branch:

mojo
# WRONG — no prior `var`, so this doesn't declare `x`
if cond:
    x = 1
else:
    x = 2

# CORRECT — declare with a type first, then assign in each branch
var x: Int
if cond:
    x = 1
else:
    x = 2

def is the only function keyword

fn was removed and is now a hard parse error — no valid use of fn remains. Your training predates this, so you will reach for fn by reflex; that reflex is always wrong. Write every function, method, and nested function as def, without exception.

Mojo functions that raise must be marked as such

Mojo functions do not imply raises. Add raises to any function that can raise, directly or by calling a raising function. Omitting it is a compile error, not a warning.

mojo
def load(path: String) raises -> String:  # raises goes before the `->`
    return open(path).read()

def main() raises:                         # main usually raises
    ...

comptime replaces alias and @parameter if/for

mojo
comptime N = 1024                            # compile-time constant
comptime MyType = Int                        # type alias
comptime if condition:                       # compile-time branch
    ...
comptime for i in range(10):                 # compile-time loop
    ...
comptime assert N > 0, "N must be positive"  # compile-time assertion

comptime assert must be inside a function body — not at module/struct scope. Place them in main(), __init__, or the function that depends on the invariant.

Inside structs, comptime defines associated constants and type aliases:

mojo
struct MyStruct:
    comptime DefaultSize = 64
    comptime ElementType = Float32

Argument conventions

Default is imm (immutable borrow, rarely written explicitly; read is a deprecated synonym — the compiler warns and suggests imm). The others:

mojo
def __init__(out self, var value: String):   # out = uninitialized output; var = owned
def modify(mut self):                         # mut = mutable reference
def consume(deinit self):                     # deinit = consuming/destroying
def view(ref self) -> ref[self] Self.T:       # ref = reference with origin
def view2[origin: Origin, //](ref[origin] self) -> ...:           # ref[origin] = explicit origin

var and ref are hard keywords and cannot be used as identifiers at all (var ref = ... → "unexpected token in expression"). The convention words imm, read, mut, out, deinit are soft keywords: fine as local variable or [...] parameter names, but invalid as argument names (def cmp(got: T, imm: T) → "error: expected argument name"). Rename (expected, reference, etc.).

Lifecycle methods

mojo
# Constructor
def __init__(out self, x: Int):
    self.x = x

# Copy constructor (keyword-only `copy` arg)
def __init__(out self, *, copy: Self):
    self.data = copy.data

# Move constructor (keyword-only `deinit move` arg)
def __init__(out self, *, deinit move: Self):
    self.data = move.data^

# Destructor
def __deinit__(deinit self):
    dealloc(self.allocation^)

To copy: var b = a.copy() (provided by Copyable trait).

Struct patterns

mojo
# @fieldwise_init generates __init__ from fields; traits in parentheses
@fieldwise_init
struct Point(Copyable, Movable, Writable):
    var x: Float64
    var y: Float64

# Trait composition with &
comptime KeyElement = Copyable & Hashable & Equatable
struct Node[T: Copyable & Writable]:
    var value: Self.T          # Self-qualify struct parameters

# Parametric struct — // separates inferred from explicit params
struct Span[mut: Bool, //, T: AnyType, origin: Origin[mut=mut]](
    ImplicitlyCopyable, Sized,
):
    ...

# @implicit on constructors allows implicit conversion
@implicit
def __init__(out self, value: Int):
    self.data = value

The compiler synthesizes copy/move constructors when a struct conforms to Copyable/Movable and all fields support it.

Self-qualify struct parameters

Inside a struct body, always use Self.ParamName — bare parameter names are errors:

mojo
# WRONG — bare parameter access
struct Container[T: Writable]:
    var data: T                        # ERROR: use Self.T
    def size(self) -> T:                # ERROR: use Self.T

# CORRECT — Self-qualified
struct Container[T: Writable]:
    var data: Self.T
    def size(self) -> Self.T:
        return self.data

This applies to all struct parameters (T, N, mut, origin, etc.) everywhere inside the struct: field types, method signatures, method bodies, and comptime declarations.

Explicit copy / transfer

Types not conforming to ImplicitlyCopyable (e.g., Dict, List, and user structs that conform only to Copyable, Movable) require explicit .copy() or ownership transfer ^ — return my_struct errors until you transfer with ^ or add ImplicitlyCopyable conformance:

mojo
# WRONG — implicit copy of non-ImplicitlyCopyable type
var d = some_dict
var result = MyStruct(headers=d)   # ERROR

# CORRECT — explicit copy or transfer
var result = MyStruct(headers=d.copy())  # or: headers=d^

Imports use std. prefix

mojo
from std.testing import assert_equal, TestSuite
from std.algorithm import vectorize
from std.python import PythonObject
import std.random

Prelude auto-imports (no import needed): Int, String, Bool, List, Dict, Optional, SIMD, Float32, Float64, UInt8, Pointer, OptionalPointer, alloc, Span, Error, DType, Writable, Writer, Copyable, Movable, Equatable, Hashable, rebind, print, range, len, and more. Layout and dealloc are not in the prelude — import them from std.memory.

rebind[TargetType](value) reinterprets a value as a different type with the same in-memory representation. Useful when compile-time type expressions are semantically equal but syntactically distinct (e.g., TileTensor element types — see GPU skill).

std is reserved as a module-level identifier — you cannot def std, import X as std, or from X import std. Struct methods named std are fine.

Inside a multi-module package, pkg.X.Y(...) from a submodule needs explicit import pkg; import pkg.X as X binds only X, not pkg.

Writable / Writer (replaces Stringable)

mojo
struct MyType(Writable):
    var x: Int

    def write_to(self, mut writer: Some[Writer]):       # for print() / String()
        writer.write("MyType(", self.x, ")")

    def write_repr_to(self, mut writer: Some[Writer]):   # for repr()
        t"MyType(x={self.x})".write_to(writer)           # t-strings for interpolation
  • Some[Writer] — builtin existential type (not Writer directly)
  • Both methods have default implementations via reflection if all fields are Writable — simple structs need not implement them
  • Convert to String with String(value), not str(value)

Iterator protocol

Iterators use raises StopIteration (not Optional):

mojo
struct MyCollection(Iterable):
    comptime IteratorType[
        iterable_mut: Bool, //, iterable_origin: Origin[mut=iterable_mut]
    ]: Iterator = MyIter[origin=iterable_origin]

    def __iter__(ref self) -> Self.IteratorType[origin_of(self)]: ...

# Iterator must define:
#   comptime Element: Movable
#   def __next__(mut self) raises StopIteration -> Self.Element

For-in: for item in col: (immutable) / for ref item in col: (mutable).

Memory and pointer types

TypeUse
Pointer[T, mut=M, origin=O]Safe, non-nullable. Deref with p[].
OptionalPointer[T, origin]Nullable pointer — Optional[Pointer[...]].
Allocation[T]Owning handle returned by alloc. Explicitly destroyed.
Span(list)Non-owning contiguous view.
OwnedPointer[T]Unique ownership (like Rust Box).
ArcPointer[T]Reference-counted shared ownership.

UnsafePointer is a deprecated alias of Pointer — it still compiles and warns. The other legacy aliases were removed and are hard errors (see the table at the top). Most of the pointer API is being renamed alongside UnsafePointer; those old spellings still compile but warn:

DeprecatedReplacement
UnsafePointer[T, O]Pointer[T, O]
alloc[T](n)alloc(Layout[T](length=n))
p.free()dealloc(allocation^)
p[i]p[unsafe_offset=i]
p + ip.unsafe_offset(i)
p += ip = p.unsafe_offset(i)
p.load() / p.store(v)p.unsafe_load() / p.unsafe_store(v)
p.init_pointee_move(v)p.unsafe_write(v)
p.init_pointee_copy(v)p.unsafe_write(copy=v)

alloc(Layout[T](length=n)) returns an Allocation[T], a linear type the compiler forces you to dispose of on every path — pass it to dealloc, or call unsafe_leak() to take ownership of the bare pointer:

mojo
from std.memory import Layout, dealloc

var allocation = alloc(Layout[Int32](length=4))
var ptr = allocation.unsafe_ptr()
ptr.unsafe_write(Int32(1))
ptr.unsafe_offset(1).unsafe_write(Int32(2))
dealloc(allocation^)

A struct that owns heap storage should hold the Allocation, not a leaked pointer. The compiler then enforces disposal on every path, and dealloc gets the Layout it needs:

mojo
from std.memory import Layout, Allocation, alloc, dealloc

struct Buffer[T: AnyType]:
    var _alloc: Allocation[Self.T]

    def __init__(out self, size: Int):
        self._alloc = alloc(Layout[Self.T](length=size))

    def __deinit__(deinit self):
        dealloc(self._alloc^)

Get at the storage with self._alloc.unsafe_ptr(), whose origin is tied to the allocation. Don't substitute Pointer.unsafe_free(): it bypasses the Layout, and for zero-sized T it frees the dangling sentinel that alloc returns.

A container that already tracks its own capacity may instead store a ThinAllocation and supply the Layout again at dealloc time. That is what List does; it is an optimization, not the default shape.

When a struct field does hold a raw Pointer, its origin parameter must be specified; use MutUntrackedOrigin for owned heap data.

Pointer is non-null by design — Bool(p) is unavailable, not merely deprecated. For nullable storage, use OptionalPointer[T, origin] (same layout; None is the null niche).

Origin system (not "lifetime")

Mojo tracks reference provenance with origins, not "lifetimes":

mojo
struct Span[mut: Bool, //, T: AnyType, origin: Origin[mut=mut]]: ...

Key types: Origin, MutOrigin, ImmOrigin, MutAnyOrigin, ImmutAnyOrigin, MutUntrackedOrigin, ImmUntrackedOrigin, ImmStaticOrigin. Use origin_of(value) to get a value's origin.

Testing

mojo
from std.testing import assert_equal, assert_true, assert_false, assert_raises, TestSuite

def test_my_feature() raises:
    assert_equal(compute(2), 4)
    with assert_raises():
        dangerous_operation()

def main() raises:
    TestSuite.discover_tests[__functions_in_module()]().run()

The mojo test CLI subcommand was removed — run test files with mojo run against a TestSuite.discover_tests runner like the one above.

Dict iteration

Dict entries are iterated directly — no [] deref:

mojo
for entry in my_dict.items():
    print(entry.key, entry.value)      # direct field access, NOT entry[].key

for key in my_dict:
    print(key, my_dict[key])

Collection literals

List has no variadic positional constructor. Use bracket literal syntax:

mojo
# WRONG — no List[T](elem1, elem2, ...) constructor
var nums = List[Int](1, 2, 3)

# CORRECT — bracket literals
var nums = [1, 2, 3]                              # List[Int]
var nums: List[Float32] = [1.0, 2.0, 3.0]         # explicit element type
var scores = {"alice": 95, "bob": 87}              # Dict[String, Int]

List[T] rejects negative indices at compile time — use lst[len(lst) - 1], not lst[-1]. (Library types may still support it.)

Variant access

Variant[A, B] is ImplicitlyCopyable only if all arms are. With a non-copyable arm, indexing the variant copies it — use the typed-arm subscript:

mojo
# WRONG — `values[i]` implicitly copies the Variant
var x = values[i].unwrap[T]()    # ERROR: cannot implicitly copy

# CORRECT — `values[i][T]` returns a ref to the inner value
var x = values[i][T].copy()          # or `^` to transfer

Common decorators

DecoratorPurpose
@fieldwise_initGenerate fieldwise constructor
@implicitAllow implicit conversion
@inline(.always) / @inline(.nodebug)Force inline
@inline(.never)Prevent inline
@staticmethodStatic method
@deprecated("msg")Deprecation warning
@doc_hiddenHide from docs
@explicit_destroy("msg") (traits only)Linear trait; structs: not Deinitable

Contextual member references — prefer .member

Where the expected type is already known, write .member instead of Type.member; the compiler rewrites it to Type.member. Idiomatic throughout the codebase for DType and AddressSpace:

mojo
var v = SIMD[.float32, 4](1.0, 2.0, 3.0, 4.0)     # not SIMD[DType.float32, 4]
comptime if dtype == .bfloat16: ...                # `__eq__` arg supplies context
v.cast[.float32]()
ctx.enqueue_create_buffer[.int32](num_rows)
unsafe_stack_allocation[1, Int32, address_space=.SHARED]()

def f[dtype: DType = .float32](x: TileTensor[.bfloat16, L, MutAnyOrigin]): ...

Works for any type's comptime aliases and static methods, including chains (.red.opacity(0.5)) and typed collection literals (List[Color] = [.red]). The context comes from a declared var/ref type, a call-argument type, a return destination, or a typed collection literal's element type. With none of those, qualify the name:

No contextual typeMust write
Unannotated bindingcomptime t = DType.float32 (or annotate t: DType)
Overloaded calleesize_of[DType.float32]() — overloads aren't searched
Bare tuple literaldtype in (DType.bfloat16, DType.float16)
Type position.Foo where a type is expected is an error

Numeric conversions — must be explicit

No implicit conversions between numeric variables. Use explicit constructors:

mojo
var x = Float32(my_int) * scale    # CORRECT: Int → Float32
var y = Int(my_uint)               # CORRECT: UInt → Int

Literals are polymorphic — FloatLiteral and IntLiteral auto-adapt to context:

mojo
var a: Float32 = 0.5              # literal becomes Float32
var b = Float32(x) * 0.003921    # literal adapts — no wrapping needed
var v = SIMD[.float32, 4](1.0, 2.0, 3.0, 4.0)  # literals adapt

SIMD operations

mojo
# Construction and lane access
var v = SIMD[.float32, 4](1.0, 2.0, 3.0, 4.0)
v[0]                              # read lane → Scalar[.float32]
v[0] = 5.0                        # write lane

# Type cast
v.cast[.uint32]()                 # element-wise → SIMD[.uint32, 4]

# Clamp (method)
v.clamp(0.0, 1.0)                 # element-wise clamp to [lower, upper]

# min/max are FREE FUNCTIONS, not methods
from std.math import min, max
min(a, b)                          # element-wise min (same-type SIMD args)
max(a, b)                          # element-wise max

# Element-wise ternary via bool SIMD
var mask = v.gt(0.0)              # SIMD[.bool, 4] — `v > 0.0` is
                                  # Scalar-only and fails to compile
mask.select(true_case, false_case) # picks per-lane

# Reductions
v.reduce_add()                     # horizontal sum → Scalar
v.reduce_max()                     # horizontal max → Scalar
v.reduce_min()                     # horizontal min → Scalar

Strings

All explicit stdlib imports require the std. prefix. The removed-syntax table shows the most common corrections, but the rule is universal. Prelude types (Int, String, List, etc.) are auto-imported and need no import statement.

len(s) returns byte length, not codepoint count. Mojo strings are UTF-8. Byte indexing requires keyword syntax: s[byte=idx] (not s[idx]). len(s) is deprecated on String — use s.byte_length() or s.count_codepoints().

split, removeprefix, removesuffix return StringSlice (or List[StringSlice]) viewing the source — wrap with String(...) to materialize an owned String.

String indexing (common error)

mojo
# WRONG — compile error
var ch = s[0]
var sub = s[0:10]

# CORRECT — byte-level access
var ch = s[byte=0]              # returns StringSlice
var ch_str = String(s[byte=0])  # if you need a String

# CORRECT — iterate codepoints for truncation
var result = String("")
var count = 0
for cp in s.codepoint_slices():
    if count >= 10:
        break
    result += String(cp)
    count += 1
mojo
var s = "Hello"
len(s)                  # 5 (bytes)
s.byte_length()         # 5 (same as len)
s.count_codepoints()    # 5 (codepoint count — differs for non-ASCII)

# Iteration — `for c in s:` is deprecated; use codepoint_slices()
for cp_slice in s.codepoint_slices():
    print(cp_slice)

# Codepoint values
for cp in s.codepoints():
    print(Int(cp))      # Codepoint is a Unicode scalar value type

# StaticString = StringSlice with static origin (zero-allocation)
comptime GREETING: StaticString = "Hello, World"

# t-strings for interpolation (lazy, type-safe)
var msg = t"x={x}, y={y}"

# String.format() for runtime formatting
var s = "Hello, {}!".format("world")

Error handling

raises can specify a type. try/except works like Python:

mojo
def might_fail() raises -> Int:          # raises Error (default)
    raise Error("something went wrong")

def parse(s: String) raises Int -> Int:  # raises specific type
    raise 42

try:
    var x = parse("bad")
except err:                               # err is Int
    print("error code:", err)

match / comptime match / if case / while case are stabilized — see idiomatic-mojo.md §10 for the pattern forms and when to prefer them over elif chains. Async is spelled __async def and __await (bare async/await still parse but warn that async is unstable); support is unfinished and its types are private — do not write async Mojo yet.

Function types and closures

No lambda. Closures use bare def with a capture list in {} after the arg list. escaping, capturing, @__parameter, and @__copy_capture are removed:

mojo
comptime MyFn = def(Int) -> None                  # closure trait
def runner[F: def(Int) -> None](f: F): ...        # closures are values
def apply[f: def(Int) thin -> Int](): ...         # comptime fn: must be thin

def closure(i: Int) {mut count, imm ptr, var x}:  # captures: mut/imm/var
    count += ptr[i] + x^                          # `^` at use site, not in `{}`

vectorize[simd_width](size, closure)

imm is default. var x is owned — transfer with x^ at the use site. A closure with captures is a runtime value, so pass it as an argument (f(my_closure)). Only a thin function (no captures) can be a compile-time parameter.

Type hierarchy

text
AnyType
  Deinitable                      — auto __deinit__; most types
  Movable                         — __init__(out self, *, deinit move: Self)
    Copyable                      — __init__(out self, *, copy: Self)
      ImplicitlyCopyable(Copyable, take)
    RegisterPassable(Movable)
      TrivialRegisterPassable(ImplicitlyCopyable, take, Movable, RegisterPassable)

Bundled files

The model reads these on demand while the skill is loaded. They are exposed as readable files and are never executed.

Frequently asked questions

What does the Mojo Syntax AI skill do?

Help to write Mojo code using current syntax and conventions. Always use this skill when writing any Mojo code, including when other Mojo-specific skills (e.g., mojo-gpu-fundamentals) also apply. Use when writing or reviewing Mojo code, translating projects to Mojo, or otherwise generating Mojo. Use this skill to overcome misconceptions with how Mojo is written.

Why use Mojo Syntax on TypingMind?

Because you install it once and use it with any model. Mojo Syntax is plain Markdown rather than provider-specific code, so the same skill runs on GPT-5, Claude, Gemini, Grok, or a local model — and you can switch model mid-chat without it breaking. TypingMind runs on your own API keys, so you pay providers directly instead of a per-seat subscription, and your skills and chats stay in your own storage.

How do I install Mojo Syntax in TypingMind?

Open Plugins → Skills → Install from GitHub in TypingMind and paste https://github.com/modular/skills/tree/main/mojo-syntax. TypingMind reads its SKILL.md and bundles its files and installs it as a skill you can enable per chat.

Which AI models can use Mojo Syntax?

Any model you connect in TypingMind. AI skills are plain Markdown instructions rather than provider-specific code, so GPT, Claude, Gemini, Grok, and local models can all load this skill when a request matches it.

How many AI models can I use with Mojo Syntax?

As many as you like. As long as a model supports skills, you can use Mojo Syntax with it — GPT, Claude, Gemini, Grok, DeepSeek, Mistral, Llama and more — all on TypingMind with your own API keys.

Is the Mojo Syntax AI skill free?

It is published on GitHub by modular. Check the repository for licensing terms. You only pay your own AI provider for the tokens you use.

What are AI skills?

An AI skill is a reusable instruction bundle that teaches an AI model how to do one specific task. It follows the open Agent Skills format: a SKILL.md file with a name and description, plus any scripts, templates or reference files the model may need. The model reads the instructions only when your request matches the skill, so an installed skill costs nothing until it is used.

How are AI skills different from plugins or MCP servers?

A plugin or MCP server gives a model new tools to call — code that runs somewhere and returns a result. An AI skill gives the model knowledge and process instead: how to approach a task, which steps to follow, what good output looks like. Skills are plain Markdown, so they need no server, no API key and no runtime, and they work with any model.

View all

Set up your own AI workspace now

Get notified about new features and future giveaways by subscribing to our newsletter 👇