Membrane Framework logo

Membrane Framework

OrganizationPopular
membraneframework
membrane-framework

Work with the Membrane multimedia streaming framework in Elixir. Use this skill whenever the user is building or debugging Membrane pipelines, writing custom Elements, Bins, or Filters, connecting pads, implementing callbacks, handling stream formats or EOS, or asking about Membrane architecture. Trigger on any mention of membrane_core, membrane, membrane framework, Membrane.Pipeline, Membrane.Sink, Membrane.Source, Membrane.Filter, Membrane.Endpoint, Membrane.Bin, Membrane.Pad, or multimedia streaming in an Elixir context — even if the user doesn't say "Membrane" explicitly but is clearly working on this codebase.

Overview

Publishermembraneframework
Repositorymembrane_core
Skill namemembrane-framework
Stars
1.5K
Forks
46
Bundled files
Instructions only
LicenseApache-2.0
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.

  • Self-contained

    Everything the model needs lives in the instructions — no extra files to sync.

  • Open source

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

Installation

Install the Membrane Framework 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.
  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/membraneframework/membrane_core.git /tmp/membrane_core
mkdir -p .claude/skills
cp -r /tmp/membrane_core/skills/membrane-framework .claude/skills/membrane-framework
Restart Claude Code after copying so it picks up the new skill.

Use it in TypingMind

Enable Membrane Framework 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 Membrane Framework 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 Membrane Framework 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.

Membrane Framework

Package: membrane_core ~> 1.3 | Docs: https://hexdocs.pm/membrane_core/ | Module index: https://hexdocs.pm/membrane_core/llms.txt | Demos: https://github.com/membraneframework/membrane_demo | All packages: packages_list.md

How to Approach Tasks

  • New component — before writing a new element, check packages_list.md to see if it already exists in an existing plugin; if not, identify subtype (Source/Filter/Sink/Endpoint/Bin), define pads, implement required callbacks (handle_buffer/4 for filters/sinks, handle_demand/5 for manual-flow sources)
  • Generating boilerplate — use mix membrane.gen.filter MyApp.MyFilter, mix membrane.gen.source, mix membrane.gen.sink, mix membrane.gen.endpoint, mix membrane.gen.bin, mix membrane.gen.pipeline instead of writing component skeletons by hand
  • Choosing element subtype — prefer Filter for transformations (has sensible defaults for stream_format forwarding); use Endpoint only when output is unrelated to input (e.g. a UDP Endpoint); use Source/Sink for pure producers/consumers
  • Flow control — default to :auto on all pads; only use :manual when you need fine-grained backpressure control; Almost the only use case of :push are output pads of Sources/Endpoints that cannot control when they produce data, e.g. UDP Source/Endpoint.
  • Pipeline topology — use the ChildrenSpec DSL (child/2, get_child/1, via_in/2, via_out/2) (more info: Membrane.ChildrenSpec)
  • Static vs dynamic topology — return spec: from handle_init/2 for static pipelines; return additional spec: actions from any callback (e.g. handle_child_notification/4) to grow the topology at runtime
  • Naming children — use atoms (:source) for singletons, tuples ({:decoder, track_id}) for multi-instance children of the same type
  • Detecting pipeline completion — implement handle_element_end_of_stream/4 in the pipeline to know when a sink's input pad received EOS; then return {[terminate: :normal], state} (doesn't work if sink is a Membrane.Bin - then expect a custom message from the bin in handle_child_notification callback instead, if the bin sends it)
  • Dynamic tracks (demuxers, variable inputs) — use the Dynamic Pads Pattern below
  • Crash isolation — group children with {spec, group: <name>, crash_group_mode: :temporary}; handle recovery in handle_crash_group_down/3; see Crash Groups guide
  • Inserting debug probes — add child(:probe, %Membrane.Debug.Filter{handle_buffer: &IO.inspect(&1, label: :buffer)}) between any two elements to log buffers without changing pipeline logic. You can use different logging functions than IO.inspect/2. More info: Membrane.Debug.Filter.
  • Linking children - linked children pads accepted formats must have non-empty intersections
  • Debugging — check pad accepted_format compatibility
  • Callback context — every callback receives ctx; key fields: ctx.children, ctx.pads, ctx.playback; crash callbacks also have ctx.crash_initiator, ctx.exit_reason, ctx.group_name; see Pipeline.CallbackContext, Bin.CallbackContext, Element.CallbackContext
  • Logging — utilize Membrane.Logger instead of Logger in Membrane components; it prepends component path and name to log messages. Requires require Membrane.Logger in the module before calling any logging functions.
  • Use mix hex.info <plugin name> when you need to check the newest version of a plugin
  • Search for appropriate plugins in packages_list.md before writing one
  • Check input and output pad definitions of elements in deps/ (use cat <filename> | grep def_input_pad and cat <filename> | grep def_output pad) to make sure output pad's accepted_stream_format is compatible with accepted_stream_format of the input pad which it is linked to.
  • If the accepted_stream_format doesn't match, search for an element which can act as an adapter
  • When constructing Membrane Pipeline, lean towards using most powerful Membrane Components, which are Boombox.Bin and Membrane.Transcoder, instead of using many smaller plugins

Common Pitfalls

  • Modifying code in deps/ directory - never do that
  • Using child/2 for an already-spawned childchild/2 always spawns a new process; use get_child/1 to reference an existing one; duplicating a name raises an error
  • Linking static pads outside the spawning spec — static pads must be linked in the same spec that spawns the component; linking them later raises a LinkError
  • Not wiring a dynamic bin pad in handle_pad_added/3 — a dynamic bin input pad must be connected to an internal child within 5 seconds or a LinkError is raised
  • Heavy work in handle_init/2handle_init is synchronous and blocks the parent; move file I/O, network connections, etc. to handle_setup/2
  • Producing data before handle_playing/2 — pads are not ready until :playing; don't send buffers from handle_setup/2
  • Returning non-spec actions from handle_init/2 — we recommend to return only :spec action from this callback.

Architecture

Pipeline
├── Element (Source/Filter/Sink/Endpoint)  ← leaf, processes data
├── Bin                                     ← dual role: parent (has children) + child (has pads)
│   ├── Element
│   └── Bin                                ← bins nest arbitrarily deep
└── ...
TypeParentChildHas Pads
Pipelineyesnono
Binyesyesyes
Elementnoyesyes

Element subtypes: Source (output only) · Filter (in + out, output is transformed input) · Sink (input only) · Endpoint (in + out, but output might be not related to input)


Pads

Defined on Elements and Bins (not Pipelines) using def_input_pad/2 (Membrane.Element.WithInputPads) and def_output_pad/2 (Membrane.Element.WithOutputPads):

elixir
def_input_pad :input, accepted_format: _any
def_output_pad :output, accepted_format: Membrane.RawAudio, flow_control: :auto
  • Availability: :always (static, one instance, referenced by atom) or :on_request (dynamic, reference via Pad.ref(:name, id))
  • Flow control: :auto (framework manages demand — preferred), :manual (explicit via :demand/:redemand), :push (no demand, risk of overflow)
  • One input pad ↔ one output pad only; pads must have compatible accepted_format
  • Default pad names :input/:output allow omitting via_in/via_out in specs
  • accepted_format matching syntax: _any (accept anything) · Membrane.RawAudio (any struct of that type) · %Membrane.RawAudio{channels: 2} (match specific fields) · %Membrane.RemoteStream{} (unknown/unparsed stream). any_of(pattern1, pattern2, ...) matches if any pattern matches.
  • Full pads guide (static vs dynamic, bin pads, lifecycle): Everything about pads

Component Lifecycle

handle_init/2      sync, blocks parent — parse opts, return initial spec
handle_setup/2     async — heavy init (open files, connect services)
                   return {[setup: :incomplete], state} to delay :playing
handle_pad_added/3 fires for dynamic pads linked in the same spec
handle_playing/2   component is ready — start producing/consuming data

All components spawned in the same :spec action enter :playing together (they synchronize to the slowest setup). Elements and Bins wait for their parent before handle_playing/2.

Stream format and EOS rules (critical for filter authors):

  • A source/filter must send {:stream_format, {pad, format}} before the first buffer on each output pad, or downstream elements crash
  • The default handle_stream_format/4 in filters forwards the format downstream — if you override it, you must forward manually or return the {:stream_format, ...} action yourself
  • The default handle_end_of_stream/3 in filters forwards EOS downstream — overriding without forwarding will stall the pipeline

Full lifecycle guide: Lifecycle of Membrane Components


Pipeline & Bin DSL (ChildrenSpec)

elixir
# Linear chain — child/2 spawns a new named child
child(:source, %Membrane.File.Source{location: "input.mp4"})
|> child(:filter, MyFilter)
|> child(:sink, %Membrane.File.Sink{location: "out.raw"})

# Explicit pad names (required for non-default names or dynamic pads)
get_child(:demuxer)
|> via_out(Pad.ref(:output, track_id))
|> via_in(:video_input)
|> child(:decoder, Membrane.H264.FFmpeg.Decoder)

# Link to an already-existing child
get_child(:existing_filter) |> child(:new_sink, MySink)

# Inside a Bin — bin_input/bin_output connect the bin's own pads to internal children
bin_input(:input) |> child(:filter, MyFilter) |> bin_output(:output)

# Crash group — all children in the spec share the group; a crash in any terminates all
{child(:source, Source) |> child(:sink, Sink), group: :my_group, crash_group_mode: :temporary}

Bin pad wiring rules:

  • bin_input(pad_ref) / bin_output(pad_ref) are the interior side of the bin's own pads
  • Dynamic bin input pads must be wired inside handle_pad_added/3 within 5 seconds or a LinkError is raised
  • Linking to a bin actually links directly to the inner component (no extra message hop)

Dynamic Pads Pattern

The standard approach for variable-track streams (e.g. MP4 demuxers):

elixir
# 1. Spawn source + demuxer; demuxer hasn't identified tracks yet
def handle_init(_ctx, state) do
  {[spec: child(:source, Source) |> child(:demuxer, Demuxer)], state}
end

# 2. Demuxer notifies parent once tracks are known
def handle_child_notification({:new_tracks, tracks}, :demuxer, _ctx, state) do
  spec = Enum.map(tracks, fn {id, _fmt} ->
    get_child(:demuxer)
    |> via_out(Pad.ref(:output, id))
    |> child({:decoder, id}, Decoder)
    |> child({:sink, id}, Sink)
  end)
  {[spec: spec], state}
end

Built-in Utility Elements

ModulePurpose
Membrane.FunnelMultiple inputs → one output
Membrane.TeeOne input → multiple outputs
Membrane.ConnectorConnect dynamic pads with internal buffering
Membrane.Testing.SourceInject buffers into a pipeline in tests
Membrane.Testing.SinkCapture and assert on buffers in tests
Membrane.Debug.FilterLog/inspect buffers flowing through pipeline
Membrane.Debug.SinkLog/inspect buffers at pipeline end
Membrane.FilterAggregatorIt is deprecated, just don't use it

Testing

elixir
import Membrane.ChildrenSpec
import Membrane.Testing.Assertions
alias Membrane.Testing

pipeline = Testing.Pipeline.start_link_supervised!(spec: [
  child(:source, %Testing.Source{output: [<<1, 2, 3>>, <<4, 5, 6>>]})
  |> child(:sink, Testing.Sink)
])

assert_sink_buffer(pipeline, :sink, %Membrane.Buffer{payload: <<1, 2, 3>>})
assert_start_of_stream(pipeline, :sink)
assert_end_of_stream(pipeline, :sink)
  • Testing.DynamicSource — like Testing.Source but with a dynamic output pad

Timing

All timestamps are Membrane.Time.t(). It is integer nanoseconds under the hood, but don't use this information, because it is a part of the private API. However, keep in mind you can perform operations on Membrane.Time.t() using + or - operators. Helpers: Membrane.Time.seconds/1, Membrane.Time.milliseconds/1, Membrane.Time.microseconds/1, etc. Timers started with :start_timer action fire handle_tick/3.

More info: Membrane.Time, Timestamps guide.


Actions

Actions are returned from callbacks as {[action_list], state}. Full reference by component type:


Key Source Files

ModulePurpose
Membrane.PipelinePipeline behaviour & all callbacks
Membrane.Pipeline.ActionPipeline action type specs
Membrane.BinBin behaviour & all callbacks
Membrane.Bin.ActionBin action type specs
Membrane.Element.BaseShared element callbacks
Membrane.Element.WithInputPadshandle_buffer/4, handle_stream_format/4, handle_end_of_stream/3
Membrane.Element.WithOutputPadshandle_demand/5
Membrane.Element.ActionElement action type specs
Membrane.PadPad definitions, Pad.ref/2
Membrane.BufferBuffer struct
Membrane.ChildrenSpecTopology DSL

Callback Reference

Callbacks are documented in the relevant behaviour modules:

Frequently asked questions

What does the Membrane Framework AI skill do?

Work with the Membrane multimedia streaming framework in Elixir. Use this skill whenever the user is building or debugging Membrane pipelines, writing custom Elements, Bins, or Filters, connecting pads, implementing callbacks, handling stream formats or EOS, or asking about Membrane architecture. Trigger on any mention of membrane_core, membrane, membrane framework, Membrane.Pipeline, Membrane.Sink, Membrane.Source, Membrane.Filter, Membrane.Endpoint, Membrane.Bin, Membrane.Pad, or multimedia streaming in an Elixir context — even if the user doesn't say "Membrane" explicitly but is clearly...

Why use Membrane Framework on TypingMind?

Because you install it once and use it with any model. Membrane Framework 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 Membrane Framework in TypingMind?

Open Plugins → Skills → Install from GitHub in TypingMind and paste https://github.com/membraneframework/membrane_core/tree/master/skills/membrane-framework. TypingMind reads its SKILL.md and installs it as a skill you can enable per chat.

Which AI models can use Membrane Framework?

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 Membrane Framework?

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

Is the Membrane Framework AI skill free?

Yes. It is published on GitHub by membraneframework under the Apache-2.0 license. 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 👇