Stats
Actions
Tags
From puredata
Use when the user wants to build, modify, or interact with Pure Data (Pd) patches — including sound synthesis (sine, FM, AM, subtractive, granular), audio routing, sequencing, control-rate logic, MIDI processing, or running Python code inside Pd via py4pd. Also use for help with FUDI, dynamic patching, [netreceive] / [netsend], or live message routing in a running Pd patch.
How this skill is triggered — by the user, by Claude, or both
Slash command
/puredata:puredataThe summary Claude sees in its skill listing — used to decide when to auto-load this skill
You have access to the **puredata** MCP server. It speaks Pd's native FUDI protocol over TCP to a `[netreceive 3000]` in a host patch the user is expected to keep open. Through the server's tools you can create objects, wire them together, toggle audio DSP, drive named receivers, and write Python scripts that become live `[py4pd]` objects on the canvas — all in one round-trip per call.
You have access to the puredata MCP server. It speaks Pd's native FUDI protocol over TCP to a [netreceive 3000] in a host patch the user is expected to keep open. Through the server's tools you can create objects, wire them together, toggle audio DSP, drive named receivers, and write Python scripts that become live [py4pd] objects on the canvas — all in one round-trip per call.
Call pd_init before any other tool. Every other tool in this MCP refuses to run until you do. pd_init returns the technical contract (wire model, object-id system, gotchas of Pd vanilla, cookbook). Read it carefully on the first call of every session — it is the authoritative source on conventions. This SKILL.md just helps you decide when to reach for the MCP; pd_init tells you how to use it.
Look for any of these intents:
osc~, dac~, metro, r, s, *~, etc.)If the user is just talking about Pd conceptually with no patch-building intent (e.g., "what is Pd?", "compare Max vs Pd"), you don't need to invoke the MCP — answer from general knowledge.
Before any MCP call can land:
${CLAUDE_PLUGIN_ROOT}/pd/mcp_host.pd. If the user reports "nothing is happening" or pd_init errors with a connection failure, the most likely cause is that this patch is not open. Tell them: "Open pd/mcp_host.pd in Pure Data and keep it running — it listens on TCP port 3000."py4pd external (≥ 1.2.3) must be installed (Pd → Help → Find externals → "py4pd"), and the patch must contain [declare -lib py4pd]. If pd_create_python_object succeeds but the [<name>] box stays empty/inert inside [pd canvas], either py4pd is missing or the patch lacks the -lib py4pd declare.For a build-from-scratch session:
pd_init — read the orientation guide. Mandatory.pd_get_state — confirm the canvas is empty (or see what's already there).pd_create_object (audio/control) or pd_create_<gui> (bng, tgl, nbx, hsl/vsl, floatatom). Place them on the canvas with sensible x/y so the user can visually follow.pd_connect, using the ids returned from step 4.pd_set_dsp(on=True). Audio is loud by default — make sure there's a [*~ <small>] (e.g. 0.1–0.2) before [dac~].pd_send_message if you wired up named [r <name>] receivers.For Python-augmented patches (py4pd 1.2.3+): use pd_create_python_object to write a .pd_py file defining a pd.NewObject subclass, and instantiate [<name>] on the canvas in one call. One file = one Pd object class. To iterate on the code, call pd_update_python_script then re-create the object on the canvas (pd_clear_canvas + rebuild, or hand-delete + recreate) — py4pd 1.2.3 does not expose a per-instance reload message.
Do not use the obsolete [py4pd script function] syntax. That mode was removed in py4pd 1.2.3 and produces inert objects with no inlets/outlets. The current API is class-based: see pd_init for the canonical template (import puredata as pd + class <name>(pd.NewObject) with name = "<name>", self.inlets, self.outlets, and in_<idx>_<msgtype> handlers).
⚠️ Two known py4pd 1.2.3 gotchas to relay to the user upfront when they ask for Python objects:
Always declare inlets as pd.DATA — never pd.FLOAT, pd.SYMBOL, or pd.LIST. Those typed inlets segfault Pd at instantiation (confirmed on Windows + Python 3.14; unverified on Linux/macOS, but pd.DATA is safe everywhere). Typed handlers (in_0_float, in_0_list, etc.) still work fine when the inlet is pd.DATA — the dispatch happens internally. Output types are unaffected (use pd.FLOAT/pd.LIST/etc. freely in self.out()).
Iterating on Python code requires restarting Pd — pd_update_python_script writes the new file, but py4pd caches .pd_py in sys.modules and never re-imports. Re-creating the [<name>] object on the canvas still runs the OLD bytecode. Tell the user to close + reopen mcp_host.pd (or restart Pd entirely) to pick up edits. Don't promise hot-reload — it doesn't exist.
Important — where the .pd_py file lands. Both Python tools take an optional scripts_dir argument. On the FIRST Python call of a session, ask the user where their Pd patch lives and pass <patch-dir>/scripts (or a path they specify) as scripts_dir. Remember that value and reuse it for every subsequent Python tool call in the session. Then tell the user their patch must contain BOTH [declare -path <that-dir>] AND [declare -lib py4pd] so py4pd is loaded and can autoregister the class — the tool response repeats both requirements you should quote back to them. If you skip scripts_dir, the server writes to the plugin's bundled pd/scripts/, which is only correct when the user is running the unmodified mcp_host.pd from the plugin install dir. Do not assume this is their setup — ask.
Environment prerequisites for the Python branch (warn the user up-front if their patch is on an older Pd, or you'll waste a round-trip on a silent failure):
undefined symbol: pd_snprintf or couldn't create).libpython3.14.so.1.0; Ubuntu 24.04 ships 3.12 by default, so the user may need to install 3.14 separately).These are the bread-and-butter shapes. The full cookbook is in pd_init's response — these are reminders so you know what's possible:
[osc~ FREQ] → [*~ GAIN] → [dac~] (mono) or [dac~] with both inlets (stereo).[osc~ MOD_FREQ] → [*~ MOD_DEPTH] → carrier [+~ CARRIER_FREQ] → [osc~] → output.[noise~] or [saw~] → [lop~ CUTOFF] / [bp~] → output. Drive cutoff via [line~] for sweeps.[bang( → message [A_VAL, A_TIME( → [line~] → [*~] on the signal.[metro MS] → [counter N] → [mod K] → branching [select 0 1 2…] → triggering.[r freq] somewhere in the patch, then pd_send_message(receiver="freq", atoms=["440"]) to update it.pd_create_slider(orientation="horizontal", min=20, max=2000) → connect its outlet to whatever should respond (often via [r/s] for clarity).[noise~] straight to [dac~]. Always insert [*~ 0.1] (or smaller) before [dac~] when prototyping.pd_clear_canvas is destructive and resets ids to 0. Confirm with the user before calling it unless they just asked for a fresh start.pd_clear_canvas and rebuild is the cleanest path.[pd canvas] mid-session, ids drift — call pd_resync_index(next_index=N) to realign (the user can read N off the canvas or you ask them).If you don't remember the exact creation arguments, inlet/outlet layout, or behavior of a Pd object, do not guess. Either:
A silently-wrong patch (empty [osc~ garbage], mismatched audio/control rates) is harder to debug than a clarifying question. Pd vanilla doesn't surface those errors back to this server.
py4pd (e.g. mrpeach for OSC, iemlib, zexy). The cookbook in pd_init lists vanilla objects only.pd_init's cookbook stays in the real-time domain..pd files on disk. The server only writes to the live canvas via FUDI; it does not parse .pd source.For any of those, fall back to general knowledge and tell the user what's out of scope.
Searches MemPalace before answering questions about past work, people, projects, or prior decisions. Returns verbatim stored content instead of guessing from model memory.
Guides Payload CMS config (payload.config.ts), collections, fields, hooks, access control, APIs. Debugs validation errors, security, relationships, queries, transactions, hook behavior.
Implements vector databases with Pinecone, Weaviate, Qdrant, Milvus, pgvector for semantic search, RAG, recommendations, and similarity systems. Optimizes embeddings, indexing, and hybrid search.
npx claudepluginhub jfboisvenue/pd-mcp-server --plugin puredata