agentleFS
Sign inStart free
sandraschisandraschi/midasheng-gen-mcp/AGENTS.md

midasheng-gen-mcp

AGENTS.md2 starsChanged 3 months agoApache-2.0View on GitHub

FastMCP 3.4 server wrapping MiDashengLM-Gen (Xiaomi Research): unified 16 kHz mixed audio scene generation (speech, music, SFX, ambience) from structured text captions. LLM-driven autoregressive flow matching with a Qwen3-1.7B backbone (2.9B params total). Apache-2.0, runs locally on CUDA. - Ports: backend 11159 (REST + MCP /mcp), frontend 11160 (registered in WEBAPP_PORTS.md) - The shared FastMCP instance lives in server.py; tool modules import mcp from there (registration at import time - no import = no tool) - uv sync --extra model…

  • Where files are headingNew AGENTS.md files have more rules about keeping secrets out of the code, and fewer about code style.

    Of the AGENTS.md files we read, the 31,774 first seen in the last six months, against the older ones: the kind of rule that gained most share, and the kind that lost most. About files like this in general, not this one, and an estimate, since a kind is a pattern in the words.

  • In this repository4 other files it gives its agents.

The file

We split this AGENTS.md into the rules it gives an agent

1 rule. About 0 are common, 1 is distinctive, and 0 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈1

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 0

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    No em dashes in any file; ASCII only

    line 60

≈ “Common” is an estimate: at least 1 in 5 of the 40,246 AGENTS.md files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
AGENTS.md

AGENTS.md - midasheng-gen-mcp

Purpose

FastMCP 3.4 server wrapping MiDashengLM-Gen (Xiaomi Research): unified 16 kHz mixed audio scene generation (speech, music, SFX, ambience) from structured text captions. LLM-driven autoregressive flow matching with a Qwen3-1.7B backbone (2.9B params total). Apache-2.0, runs locally on CUDA.

Layout

src/midasheng_gen_mcp/
  server.py          FastMCP instance (shared), lifespan, FastAPI build
  main.py            Dual transport entry (stdio default, HTTP on MCP_PORT)
  config.py          Settings from MIDASHENG_* env vars
  model_manager.py   Lazy torch/transformers load, generate, state machine
  jobs.py            Async job runner + prompt builder
  db.py              SQLite scenes + jobs store (WAL)
  samples.py         Built-in example captions
  audio.py           WAV metadata + persistence helpers
  logs.py            Ring-buffer log store
  mcp/tools/         audio_scene portmanteau, help, prefab cards
  api/routes/        REST: health, dashboard, scenes, jobs, audio, tools,
                     skills, samples, logs, llm, model, capabilities
  skills/midasheng-gen/SKILL.md
webapp/              React + Vite + Tailwind SOTA dashboard (port 11160)
tests/               15 pytest tests (no torch required)

Key facts

  • Ports: backend 11159 (REST + MCP /mcp), frontend 11160 (registered in WEBAPP_PORTS.md)
  • The shared FastMCP instance lives in server.py; tool modules import mcp from there (registration at import time - no import = no tool)
  • uv sync --extra model installs torch/transformers; without it the server runs but reports not_installed state (declared, not fake)
  • Model checkpoint: mispeech/midashenglm-gen on Hugging Face (~6 GB), auto-download via start.ps1 or the download_model operation
  • Structured caption: <|caption|> <|asr|> <|speech|> <|sfx|> <|music|> <|env|> views; absent views become <|unknown|>
  • MCP generate is synchronous; the webapp uses the async job API (POST /api/generate -> job id -> GET /api/jobs/{id})

Commands

just lint         ruff check + format --check
just test         pytest (light, no torch)
just web-check    tsc --noEmit + biome check (webapp)
just ci           lint + test
just mcpb-pack    fresh-stage MCPB bundle
just model-download   installs model extra + downloads checkpoint

Conventions

  • Fleet standards: mcp-central-docs (AGENTS.md gate, SOTA, WEBAPP_PORTS)
  • No em dashes in any file; ASCII only
  • Dialogic returns: {success, message, ...}; _error_response auto-logs
  • Prefab cards for list/status surfaces (show_scene_status_card etc.)

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

Reviews can't be read right now.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

Check your setup, free