analog-rytm-agent-bridge
Click on "Install Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type
@followed by the MCP server name and your instructions, e.g., "@analog-rytm-agent-bridgeRead the current kit and summarize the kick drum settings"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
analog-rytm-agent-bridge
Let your coding agent use your Analog Rytm MkII with MCP.
Thirty-two semantic tools covering the whole instrument — patterns and p-locks, kit sounds and machines, scenes and performance macros, songs, the sample RAM, and audio capture. The agent reads the device's real state, proposes a change in musical terms, and commits it on the next beat.
Site · MCP setup · Quick start · Tools · Safety · Capabilities
claude mcp add rytm -- npm run mcpThat attaches the full tool surface to your agent against a mock device — no
hardware, no MIDI cable, nothing at risk. Ask it "what's on pattern A01?" and
it will reach for rytm_inspect_pattern.
What this is
An Analog Rytm holds twelve voices, 128 patterns, kits, sounds, scenes, performance macros, songs, and 127 sample slots of state that a person edits by hand. This bridge exposes that state to a coding agent as semantic operations, so you can say "give the snare a conditional fill every 4th pass" or "chain A01 into A02 twice and mute the toms on the last one" and have it land on the hardware, on the beat.
The vocabulary is musical intent rather than SysEx addresses, and the agent works from a compact summary of what is actually on the device rather than a guess. Capture is part of the loop too — it can play what it wrote, record it, and listen back.
It is also usable without an agent. The daemon and CLIs are ordinary tools for inspecting, validating, queueing, and applying Rytm operations.
Related MCP server: forge-mcp
The safety model
The short version: hardware has no ctrl-Z, so every persistent change is validated against decoded device state, snapshotted as raw SysEx, applied at a musical boundary, read back, and restored byte-exactly if the readback disagrees. Read the rest before pointing it at hardware.
Nothing mutates without --execute. Every hardware command runs in
validate-only mode by default: it connects, decodes real device state, checks
the operation against it, and reports what would happen. Adding --execute
is the only way to write.
Every persistent write is snapshotted first. The daemon captures the raw SysEx of each affected Pattern, Kit, Global, and Settings object before touching it, and stores it durably.
Every write is verified by readback. After applying, the daemon re-reads the object and compares it to what was asked for, canonicalizing codec-quantized values first so the comparison is real rather than cosmetic.
A failed readback rolls back automatically — restoring the original raw bytes, across multiple objects, without ever decrementing the public revision. This path is hardware-certified, not just unit tested: see docs/HARDWARE_VALIDATION_2026-07-17_COMPLETE.md.
What it will still overwrite. The bridge protects the objects it knows it is touching. It does not back up your entire device. Before first use, save your projects to +Drive and take an external backup — this is a tool that writes to a musical instrument you care about.
Realtime gestures are deliberately not persistent. Scene activation, performance macro amounts, and live parameter moves go out as transient CC/NRPN and never change the persistent revision, matching how the hardware itself treats them.
Requirements
macOS. The daemon depends on CoreMIDI and CoreAudio and does not build elsewhere.
An Analog Rytm MKII. Certified against OS 1.72; the codecs target firmware 1.70 and unknown-capability operations are gated rather than guessed.
Rust ≥ 1.89 (
File::try_lock, used for the single-instance state lock).Node ≥ 22.14 (24+ recommended). The TypeScript side runs on Node's native type-stripping. Its only runtime dependency is the official MCP SDK, used by the stdio server; the control plane, CLIs, and daemon client pull in nothing.
For sample management only: a pinned Elektroid CLI fork — see docs/HARDWARE_SETUP.md.
Quick start
1. Without hardware
git clone https://github.com/chronick/analog-rytm-agent-bridge
cd analog-rytm-agent-bridge
npm install # MCP SDK + dev-time typechecker
npm run demonpm run check runs the full gate: Node tests, TypeScript typecheck, and the
Rust daemon's cargo test.
Run the mock daemon as a long-lived process to exercise the real RPC boundary:
cargo run --manifest-path daemon/Cargo.toml -- serve --adapter mock2. With hardware
Work through docs/HARDWARE_SETUP.md first — it covers MIDI port configuration, the device settings the bridge expects, and the backup you should take before any write test.
Confirm the device is visible and identifies itself:
cd daemon
cargo run -- midi-list
cargo run -- identity
cargo run -- capture-state ../hardware/runs/baselineThen run the validation suite. Without --execute it only reads:
npm run hardware:control # validate only
npm run hardware:control -- --execute
npm run hardware:all -- --execute --phase=coreStart the hardware daemon:
cargo run --manifest-path daemon/Cargo.toml -- serve --adapter hardware --clock-source observedState lives in ~/.analog-rytm-agent-bridge/hardware-state.json by default;
--state-dir selects an isolated store. Mock and hardware modes speak the same
request/response/event protocol — see docs/DAEMON_RPC.md.
Use it from an agent
The bridge ships an MCP stdio server exposing all 32 tools to any MCP client. Start with the mock adapter — every tool answers, and no hardware is involved:
claude mcp add rytm -- npm run mcpFor the real device:
claude mcp add rytm-hw -- npm run mcp:hardwareFull setup for Claude Code, Claude Desktop, and other clients — plus flags and troubleshooting — is in docs/MCP.md.
The agent surface
Thirty-two semantic MCP tools, grouped by what they let an agent do:
Group | Tools |
Inspect |
|
Propose & commit |
|
Play |
|
Undo |
|
Samples |
|
Listen |
|
Meta |
|
rytm_describe_capabilities is the one an agent should call first: it reports
what the connected device and firmware actually support, so unsupported
operations fail as a refusal rather than a corrupted object.
Declarative projects
build:project applies a whole project — patterns, machines, sounds, scenes,
performance macros, samples — from one JSON declaration, validation-first, with
snapshot and readback:
npm run build:project -- <declaration.json> [--execute] [--auto-slots]
npm run audition:project [-- A01 B04 ...]A declaration with a samples section is preflighted against the device's RAM
inventory before anything is applied. Each declared slot must be free or
already hold that sample's own content; otherwise the run prints a conflict
report and exits non-zero with nothing applied, rather than failing late after
every kit and pattern batch has already landed. --auto-slots remaps
conflicting slots in memory (lowest free slot first, and a sample already
loaded elsewhere follows its own slot, so repeat runs are idempotent), rewrites
the sample_number p-locks and kit slot fields that referenced them, and
prints the final map as one line of JSON. P-lock references to slots the
declaration does not own are never touched.
The sounds section designs each track's kit sound — a machine selection plus
per-page parameter locks. Every field is optional. The machine is emitted
before its parameters, because setting a machine resets its page to defaults:
{
"project": "layered-kick-demo",
"patterns": [],
"sounds": {
"BD": {
"machine": "bdplastic",
"machineParams": { "tun": -14, "swt": 54, "swd": 21, "dec": 45, "tic": 32, "lev": 110 },
"filter": { "filter_type": "Pk", "resonance": 40 },
"amp": { "overdrive": 8 }
},
"BT": {
"machine": "btclassic",
"lfo": { "destination": "SampleFineTune", "waveform": "Tri", "mode": "Hold", "depth": 32 }
}
}
}Parameter names and enum casing are the daemon's — see apply_sound_parameter
in daemon/src/hardware.rs. Enum values are the CamelCase serde variants
(SampleStart, Tri), not the lowercase rytm-rs strings.
audition:project then plays each pattern from generated clock and captures a
verified bounded recording per slot.
Architecture
Coding agent / MCP host
→ MCP stdio server / TypeScript facade
→ versioned JSON-lines RPC over stdio
→ long-running Rust daemon revisions, queue, snapshots, rollback
→ CoreMIDI / SysEx / realtime MIDI
→ Analog Rytm MKII
↘ CoreAudio capture (stereo, or Overbridge multitrack)Two lanes, deliberately separate: SysEx for persistent state, realtime MIDI for gestures. Two adapters behind one protocol: a mock for development and a hardware adapter that adds a durable queue, explicit transport epochs, generated or observed MIDI clock, reconnect reconciliation, and semantic readback verification. Deltas rather than whole-project regeneration; compact state summaries rather than giant payloads.
This repo is intentionally separate from pd-agent-bridge, which is the
reference implementation of the control-plane pattern, not a dependency. They
are meant to run side by side with the coding agent as the only glue, and each
should remain useful alone.
Documentation
Doc | What's in it |
The overall design and its rationale | |
Everything implemented, and what is not | |
Read before any write test | |
The JSON-lines protocol | |
Operation and recovery sequence for an agent | |
The current control matrix | |
Evidence-driven mapping of every manual control family | |
The stereo capture contract | |
Optional synchronized multitrack capture | |
Sample identity, transfer, RAM resolution, rollback boundaries | |
Wiring the MCP server into Claude Code, Claude Desktop, or your own client | |
The | |
Second-opinion review findings and dispositions |
The dated HARDWARE_VALIDATION_*.md files are certificates: each records a
suite run against a connected device, with the device restored to its exact
pre-test state afterward.
Credits and licensing
SysEx codecs come from rytm-rs by
alisomay (MIT), pinned to a maintained fork whose changes are being prepared
for upstream — see docs/UPSTREAM.md. Sample transfer uses a
pinned fork of Elektroid by dagargo.
docs/reference/rytm.yaml is a machine-readable parameter reference distilled
from Elektron's published documentation. Its descriptions are concise
paraphrases, not copies. Elektron's manual is copyrighted and is not
redistributed here; keep your own copy if you want one.
This project is not affiliated with, endorsed by, or supported by Elektron. "Analog Rytm" and "Overbridge" are Elektron's trademarks, used here only to describe what the software interoperates with.
Licensed under the MIT License.
This server cannot be installed
Maintenance
Resources
Unclaimed servers have limited discoverability.
Looking for Admin?
If you are the server author, to access and configure the admin panel.
Related MCP Connectors
Create, co-edit, analyze, publish, and export collaborative step-sequencer sessions through MCP.
System-of-record notebook for AI coding agents: pages, datastores, tasks, skills over MCP.
Your org's AI agents, tasks, runs, search, and brain files as MCP tools and resources.
Agent-native notes, tasks, dev-docs, vaults, sync & handoffs. MCP + OpenAPI dual surface.
Related MCP Servers
- FlicenseNot gradedqualityDmaintenanceEnables AI agents to control and monitor Bitwig Studio in real-time using natural language commands through MCP and OSC.
- AlicenseNot gradedqualityBmaintenanceEnables MCP-capable agents to author, compile, run, and commit generative music E-- Recipes from a library catalog, completing the full authoring loop.Apache 2.0
- AlicenseNot gradedqualityBmaintenanceAn MCP server that exposes Ableton Live control (session state, transport, tracks, devices, clips, MIDI note editing) as tools for LLM agents, enabling natural language manipulation of a Live session.1MIT
- AlicenseAqualityBmaintenanceProvides transactional intelligence for AI agents, enabling safe tool execution with pre-flight invariant checks, sub-second filesystem snapshots/rollback, causal tracing, belief contradiction detection, and 15 native MCP tools for Claude Code.15MIT
Latest Blog Posts
- Who's Calling? MCP Hosts Are an Identity Blind Spot (And the Spec Knows It)By Om-Shree-0709 on .mcpAgent IdentityOAuth 2.1
- Your AI Chatbot Just Exposed Your CEO's Salary to an InternBy Om-Shree-0709 on .Agent IdentityMCP SecurityOAuth Delegation
- Why MCP Servers Need Execution Sandboxing (And Why Your Current Stack Isn't Enough)By Om-Shree-0709 on .Agentic AiPrompt InjectionWebAssembly
MCP directory API
We provide all the information about MCP servers via our MCP API.
curl -X GET 'https://glama.ai/api/mcp/v1/servers/chronick/analog-rytm-agent-bridge'
If you have feedback or need assistance with the MCP directory API, please join our Discord server