Files
soundcraft/docs/mcp.md
T
aiongg e82c1689b3 refactor(cli): parse soundcraft-cli arguments with clap, following RFC-0012 (#1136)
* test(cli): pin today's argument handling

Run the real binary and record how the hand-written parser behaves for every
subcommand: help and version spellings, no arguments, unknown subcommands and
flags, value flags in both spellings, missing and repeated values, positional
order, exit codes and which stream carries each message.

No behaviour change. Later parser changes update these assertions on purpose.

* refactor(cli): split main.rs into modules

Move only: args, error, output, convert, info, session, catalog, app and
mcp each get their functions and unit tests; main.rs keeps the usage doc,
help_text and the dispatch. Moved items become pub(crate) where other
modules call them. No behaviour change.

* refactor(cli): typed errors and run(args, out, err)

soundcraft-cli is now a library plus a thin binary. `soundcraft_cli::run(args,
out, err) -> i32` is the whole program; main.rs only passes the real streams
and shuts the plugin hosts down afterwards (process teardown, not part of run,
which tests call repeatedly in one process). Subcommands write to the given
writers, return `Result<(), CliError>`, and `fail` is gone. CliError wraps
engine, audio-io, io (with the path), serde_json and the control channel
errors; UsageError, ConvertError and RootFlagError cover the subcommands' own
failures. No Result<_, String> is left except where a dependency reports a
string (control-channel call, AutomationRoots::new), wrapped in a variant.

The tests that need no files, working directory, environment or stdin now call
`run` in-process (tests/it/run.rs); the rest still run the binary.

Behaviour changes:
- Error messages start with `error: ` instead of `soundcraft-cli: `.
- A script line that stops the run now prints `error: FILE:N: ID: ...` (it had
  no prefix). With --keep-going the line stays bare.
- A convert encode or write failure now names the output path.
- An empty or help-less call still prints the one-line usage to stderr without
  a prefix and exits 1; `files confined:` keeps its `soundcraft-cli:` prefix
  (it is a note, not an error).

* refactor(cli): parse arguments with clap

Replace the hand parser (arg_value, check_options, ENGINE_OPTIONS, the
lib.rs dispatch, the per-subcommand positional handling and the root
flags' consume) with clap derive types in args.rs: a Parser, a
Subcommand enum and one Args struct per subcommand, with the run/script
engine flags flattened. main passes args_os(); run takes OsString-like
arguments and renders clap's help, version and usage errors (help and
version on stdout, exit 0; usage errors on stderr, exit 2). Add
command() returning the clap definition. The usage block in the module
doc and help_text are gone: the doc comments are the help.

Behaviour changes:
- Usage errors exit 2 with clap's message and the --help hint on
  stderr, instead of exit 1 with hand-written text: unknown subcommand,
  unknown flag, missing value, missing or extra positional, a value on
  a boolean flag (--demo=1, --normalize=1).
- No arguments prints the help to stderr, exit 2 (was a one-line usage,
  exit 1).
- --help, -h and help print clap's generated help, and every subcommand
  has its own (SUB --help, SUB -h, help SUB). Before, info and script
  read --help as a file name, describe as an id, commands as a filter,
  run rejected --help and ignored -h, convert and app printed their
  usage error, mcp, parity and plugins ignored it, and app --port P -h
  called the method -h.
- --help or -h before -- shows the subcommand's help even after a
  mistake on the same line (run --bounce --help).
- --version, -V and version print "soundcraft-cli X.Y.Z" (was
  "SoundCraft CLI X.Y.Z").
- Unknown flags on info, commands, describe, app, mcp, parity and
  plugins are usage errors; they were ignored or read as a positional.
  mcp --automation-bogus exits 2 (was exit 1).
- Extra positionals are usage errors on every subcommand (were ignored).
- A single-dash word (-x) is a flag, so an unknown one is a usage
  error; a value starting with - is written --flag=-value.
- Flags may come before or between the positionals: convert IN --rate
  44100 OUT, convert --rate 44100 IN OUT and script --keep-going FILE
  parse (the flag word was taken as a file name).
- A repeated single-value flag keeps its last value (was the first):
  --rate, --bit-depth, --save, --port and the others.
- parity --write without a value is a usage error (printed the report).
- mcp --connect without a value, or followed by a flag, is a usage
  error (was exit 1, or took the flag as the port).
- mcp --demo=1 is a usage error (was ignored, without the demo).
- A root flag without a folder, with an empty one, or followed by a
  flag is a usage error with clap's message (was exit 1).
- A non-Unicode argument is a usage error where text is expected and
  accepted as a root folder (std::env::args panicked).
- The help lists --sample as --demo's alias.

Unchanged: --dither with --no-dither, bad --rate and --bit-depth values,
the roots with --connect and the SOUNDCRAFT_CONTROL_PORT fallback are
still checked after parsing (exit 1).

* refactor(cli): reject repeated flags, clap --version (photocraft#7071 decisions)

Behaviour changes:
- A flag given twice is a usage error (exit 2), in either spelling, including
  an alias pair such as --demo and --sample. Before, the last one won. Only
  run's --cmd and mcp's --automation-read-root / --automation-write-root
  repeat; they accumulate in order.
- --version and -V print the long version, "soundcraft-cli 0.13.1 (linux
  x86_64)": the version and the platform the binary was built for.
- The hidden `version` subcommand is gone; `soundcraft-cli version` is an
  unrecognized-subcommand usage error.
- --start and --end accept a negative number (`--start -1`, as the engine
  counts positions before 0); any other value starting with `-` needs `=`
  and otherwise is an unexpected-argument usage error. No flag has
  allow_hyphen_values.
- `help <subcommand>` is unchanged (tested).

* refactor(cli): typed values and flag relations

Handlers receive typed values: convert --bit-depth is a ValueEnum, --rate a
NonZeroU32, file arguments PathBuf, --start/--end a TimeArg, --cmd a parsed
(id, JSON) spec, app's JSON a serde_json::Value, app --port and mcp --connect
a validated port or host:port. Relations (--dither/--no-dither, --in/--demo,
--start/--end needing --bounce) are declared on the flags. describe ID,
commands FILTER and app METHOD refuse an empty string. parse_rate,
parse_bit_depth, parse_dither, parse_cmd_spec, time_arg and the Rate,
BitDepth, DitherConflict and CmdJson errors are gone.

Behaviour changes:
- --dither with --no-dither is a usage error (exit 2) instead of exit 1.
- --rate 0 / non-numeric and an unknown --bit-depth are usage errors (exit 2)
  with clap's wording, naming the flag. --bit-depth help lists the values.
- --in with --demo (or --sample) is a usage error; --in used to win silently.
- --start/--end without --bounce are a usage error; they used to be ignored.
- Malformed JSON in --cmd ID=JSON and in app's JSON argument is a usage error
  (exit 2) at parse time. Before, run opened the engine and executed every
  earlier --cmd (printing their results) before the bad one failed with exit 1.
- An empty --start/--end, an empty command id (--cmd =1), and an empty
  describe ID, commands FILTER or app METHOD are usage errors. An empty
  commands FILTER used to list everything.
- app --port and mcp --connect accept only a port (0-65535) or host:port;
  junk used to be accepted and failed at connect time (exit 1).
- Non-Unicode file names are accepted wherever the callee takes a Path (info
  audio files, convert, script, parity --write). Session and bounce/save paths
  go to the engine as text, so a non-Unicode one fails with exit 1, "the path
  is not valid Unicode", after any earlier --cmd ran.

* refactor(cli): shared launch flags and SOUNDCRAFT_CONTROL_PORT as clap env

soundcraft_automation::args holds the launch flags as clap Args structs:
ControlClientArgs (`app --port`, env SOUNDCRAFT_CONTROL_PORT, default
7979), ControlServerArgs (the desktop `--control`, same variable, wired
in a later step; its ControlPort value is a port, also from a loopback
host:port, or another host's host:port, which the app decides on, so one
export serves both), AutomationArgs (the two root flags, repeatable, flags
only), the ControlTarget value parser, and the blank_env_is_unset,
name_env_vars and conflict helpers. The module and its clap/thiserror
dependencies are left out of the wasm build. `app` and `mcp` flatten
them; the hand fallback to the variable in `app` is gone.

Behaviour changes:
- An empty or blank SOUNDCRAFT_CONTROL_PORT counts as unset, so `app`
  uses 7979; before, it tried to connect to `127.0.0.1:` (exit 1).
- A bad SOUNDCRAFT_CONTROL_PORT is a usage error naming the variable
  (exit 2); before, the connection failed (exit 1).
- Spaces around the `app --port` or `mcp --connect` value, or the
  variable's, are ignored; before, the connection failed (exit 1).
- An IPv6 host in `app --port`, `mcp --connect` or the variable goes in
  brackets (`[::1]:7979`); `::1:7979` is a usage error (exit 2), as the
  host and port can't be told apart; before, the connection failed
  (exit 1).
- `app --help` shows `[env: SOUNDCRAFT_CONTROL_PORT=]` and
  `[default: 7979]` instead of the hand-written default note.
- `mcp` with a root flag and --connect is clap's conflict error (exit 2,
  "the argument '...' cannot be used with '...'"), reported before any
  root folder is opened; before, a custom message with exit 1, after the
  roots were opened.

* refactor(app): move launch flag parsing into launch_args

The parsing moves out of main unchanged, into a LaunchArgs struct and a
parse function that takes SOUNDCRAFT_CONTROL_PORT as a parameter. Tests pin
the flags, the file list, ignored unknown flags and the fallback of a bad
--control to the variable. No behaviour change.

* refactor(app): parse launch flags with clap

soundcraft --help now comes from the LaunchArgs doc comments and ends with the
environment variables the app reads, and the shared ControlServerArgs supplies
--control and SOUNDCRAFT_CONTROL_PORT. Process tests cover the variable (bad
value, blank value, loopback host:port, another host).

Behaviour changes:
- An unknown flag (--bogus, -x) is a usage error on stderr, exit 2. It was
  ignored, and a single-dash word was opened as a file; after `--` it is a file.
- A bad --control (out of range, not a number, missing value) is a usage error,
  exit 2. It was ignored and the variable used instead.
- A bad SOUNDCRAFT_CONTROL_PORT is a usage error naming the variable, exit 2.
  It was ignored. An empty or blank variable still counts as unset.
- A loopback host:port (127.0.0.1, localhost or [::1]) in --control or
  SOUNDCRAFT_CONTROL_PORT gives its port, so the value `soundcraft-cli app`
  takes from the variable works for both. The control server listens on
  127.0.0.1 only, so another host in --control is a usage error. In the
  variable, which may point `soundcraft-cli app` at another machine, it is
  ignored with a log warning and the app opens without a control server.
  Both were ignored before.
- A repeated flag is a usage error (--demo twice, --control twice).
- --control=PORT and a port with surrounding spaces are accepted; a value that
  starts with `-` is not taken as the value, except a negative number, which is
  a bad port (`--control -1`).
- --help ends with an Environment section listing SOUNDCRAFT_CONTROL_PORT,
  SOUNDCRAFT_CONFIG_DIR, SOUNDCRAFT_NO_PREFS, SOUNDCRAFT_NO_SYSTEM_FONTS,
  SOUNDCRAFT_WAYLAND (Linux) and RUST_LOG.
- --help and -h print the help, exit 0. They were ignored.
- --version and -V print `soundcraft 0.13.1 (linux x86_64)`, as soundcraft-cli
  does. It printed `SoundCraft 0.13.1`.
- -V and --version are read in order with the other arguments: an earlier
  error comes first.
- macOS -psn_… process serial numbers are dropped, as before they were
  opened as files.
- A file whose name is not valid Unicode is skipped with a log warning. It
  used to panic in std::env::args.

* feat(cli): shell completions and man pages

An example, `cargo run -p soundcraft-cli --example man-and-completions -- <out-dir>`,
writes bash, zsh, fish, elvish and PowerShell completions and soundcraft-cli.1 plus a
page per subcommand, using clap_complete and clap_mangen (dev-dependencies of
soundcraft-cli only, not in the binary).

The Linux package (deb, rpm, tar.gz, AppImage) installs the pages in
/usr/share/man/man1 and the bash, zsh and fish completions in their standard
directories; the deb also links the zsh completion into Debian's
vendor-completions. A cross build runs the example through $EMULATOR and
leaves the files out without one. The FreeBSD tarball carries them in the
ports layout; a cross build builds the example for the host as well.

No change to the CLI's behaviour. A test pins the files the example writes.

* ci(cli): check the command-line code with ast-grep rules

Adds lint/ast-grep (rules for clap errors built from strings, String as an
error type, and std::env::args) and a workflow that runs the rule tests and a
scan over the CLI code through the shared craft-common-actions workflow. No
behaviour change; the scan has no hits.

* docs(cli): document soundcraft-cli and RFC-0012

Add docs/cli.md: the subcommands, exit codes, environment variables, the
desktop app's launch flags, man pages and completions, and the RFC-0012 rules
not followed yet (convert can overwrite its input; no --compact/--json).
Link it from the README and the MCP and control-protocol guides. No code change.

* feat(cli): refuse an output that replaces an input

An output that is one of the subcommand's inputs is a usage error (exit 2,
`error: …` on stderr, clap's usage line), reported while parsing, before
the engine loads or any --cmd runs. Paths are compared by file identity
(device and inode, through same-file), or, when neither exists yet, by
where they would be created, so `a.wav`, `./a.wav`, an absolute path and a
symbolic or hard link all count as the same file.

Inputs and outputs per subcommand:
- convert: input IN, output OUT. --in-place lets OUT replace IN.
- run: input --in, outputs --bounce and --save. --in-place lets --save
  replace the --in session; it requires --in. A --bounce naming the --in
  session is always refused.
- script: inputs FILE and --in, outputs --bounce and --save. --in-place
  lets --save replace the --in session; a --bounce naming it is always
  refused, and neither may ever replace FILE.
- run and script: --bounce and --save naming one file is refused with or
  without --in-place (the save would replace the bounce).
- info, commands, describe, app, mcp, plugins write no file; parity
  --write has no input. Paths inside --cmd parameters are not checked.

With --in-place, an output that is a symbolic link to the input is written
to the file the link names, so convert and --save both replace the input
and leave the link in place (a save renames a new file over its path,
which would otherwise replace the link and leave the input unchanged).

Behaviour changes:
- `convert a.wav a.wav`, `run --in s.scraft --save s.scraft`,
  `run --in s.scraft --bounce s.scraft` and the other cases above exit 2
  instead of replacing the input.
- New flag --in-place on convert, run and script.
- An existing file that is not an input is still replaced without asking.

* feat(cli): --compact for JSON output

RFC-0012 rule 1.3: JSON is pretty-printed by default and --compact prints
one line. Every pretty-printed JSON output now goes through
output::print_json.

Changed outputs:
- info <audio file>, describe, app: new --compact flag.
- run and script: new --compact flag, which requires --inspect and applies
  to the --inspect JSON only.

Unchanged on purpose:
- The `ID: JSON` result lines of run and script and the --bounce and --save
  results stay one line each. The prefix makes them text records, and the
  RFC lets a command document its own layout; scripts that parse them keep
  working.
- commands, plugins, parity and the .scraft summary of info are text and get
  no --json flag (an open question in the RFC); mcp speaks JSON-RPC lines.

Pretty output is byte-identical to before.

* fix(app): one Switch parser for boolean environment switches

SOUNDCRAFT_NO_PREFS, SOUNDCRAFT_NO_SYSTEM_FONTS and SOUNDCRAFT_WAYLAND each
counted any value, including 0, as on. They now share
soundcraft_ui_egui::switch::Switch: 1/on/true/yes is On, 0/off/false/no is
Off (case- and space-insensitive), unset or blank is Default, and any other
value is Default and logged with the variable name. The ENVIRONMENT help
text, docs/cli.md and the README say the same.

Behaviour changes (old -> new):
- SOUNDCRAFT_NO_PREFS=0 disabled preferences (any value did); now only On
  does.
- SOUNDCRAFT_NO_SYSTEM_FONTS=0 disabled system fonts (any value did); now
  only On does.
- SOUNDCRAFT_WAYLAND=0 kept native Wayland (any value did); now only On does.
- An empty value of any of the three counted as set; it now means the
  default (off).
- An unknown value of any of the three is logged.

* refactor(app): read the environment once into Env

The desktop app reads SOUNDCRAFT_CONFIG_DIR, SOUNDCRAFT_NO_PREFS,
SOUNDCRAFT_NO_SYSTEM_FONTS and SOUNDCRAFT_WAYLAND once at startup into
env::Env (typed fields, Switch for the on/off ones), after the logger is
installed so malformed switches are logged. The --help Environment
section is generated from the same VARS table, with a test that it lists
every name.

ui-egui no longer reads the app's variables: fonts::install and
definitions take system_fonts, set through SoundApp::set_system_fonts
(default on, so the web app is unchanged). The XWayland and prefs
decisions are methods on Env, tested on literals. logging::install takes
the RUST_LOG value, which stays outside Env because the logger needs it
before anything can log.

No behaviour change beyond help column spacing.

* refactor(cli): one long_version for both binaries

* refactor(audio-io): one FileFormat table for extensions

FileFormat gains the extension table every caller matches on:
from_extension / from_path (ASCII case-insensitive) over extensions,
from_write_extension over write_extensions, default_extension, name,
mime_type, the READ and WRITE lists, and readable_extensions /
writable_extensions for dialog filters, help and messages. Reading
takes every extension the decoder knows; writing takes only the names
whose readers expect what the encoder writes (no rf64 or aifc, whose
files would get plain RIFF and AIFF headers).
detect_format, bounce (mix and stems), convert, the bounce dialog, the
desktop and web file pickers, the workspace browser, the Sampler, web
downloads and the MCP bounce schema all use it; ENCODE_EXTENSIONS and
encode_format_for are gone. convert's IN/OUT and info's FILE help name
the extensions from the read and write lists.

Behaviour changes, where the lists disagreed:
- The bounce dialog's Browse result, the bounce save filter and the MCP
  bounce enum take wav, wave, bwf, aif, aiff and flac: wave and bwf are
  new to all three, and aif is new to the MCP enum. convert and
  file.bounce_mix already took these. The "supported:" list in the
  cannot-write errors is "wav, wave, bwf, aif, aiff, flac".
- file.bounce_stems reads its format name case-insensitively: `AIFF`
  or `FLAC` meant WAV before. Unknown names still mean WAV.
- The bounce dialog sends the format as its file extension (aif for
  AIFF); the engine accepts both spellings.
- File pickers, the workspace browser and the Sampler's drop and Load
  Sample filter list every extension the decoder reads: the workspace
  browser gains wave, bwf, rf64, aifc, mp1, mp2, oga, aac, mp4, adts,
  mkv and webm; the Sampler gains rf64, mp1, mp2, m4b, adts, mkv and
  webm; the desktop and web open filters gain rf64, mp1, mp2 and adts.
- Web downloads get a media type for every audio format, not only WAV,
  AIFF and FLAC.

* refactor(app): resolve the settings folder once in Env

* fix(app): keep paths as paths until the engine needs text

A path is text only where an engine command takes `"path": string`; a
name that is not valid Unicode is reported there instead of being opened
under a lossy rendering (another file, or a confusing "can't read").

- ui-egui: dialogs::unicode() and NotUnicode ("<path:?> is not valid
  Unicode"); dialogs::open_path(app, &Path), moved from the desktop app,
  is the one open-by-extension path for dropped, command-line and
  Workspace files.
- Services: the file dialogs return PathBuf (pick_save and pick_folder
  start from a &Path); callers convert where the text field or command
  takes the path.
- Sampler drops and Load Sample pass the path to mix.sampler_load
  through unicode(); the audio test checks the Path.
- Workspace panel keeps entry paths; names are shown lossily only as
  labels. workspace_dir stays a String (serialized preferences).
- soundcraft-cli: utf8() is unicode(), CliError::NonUnicodePath is
  NotUnicode; it is called only for info's session, --in, --bounce and
  --save. convert's output format test no longer treats a non-Unicode
  extension as "no extension".
- docs/cli.md: sessions and bounces the engine opens or writes must be
  valid Unicode.

Behaviour changes:
- A dropped, command-line, picked or Workspace file whose name is not
  valid Unicode is not opened: the status line (and log) says
  "<path:?> is not valid Unicode", with the bytes escaped. Dropped files
  used to be opened under a U+FFFD rendering.
- A file dialog result that is not valid Unicode leaves the path field
  unchanged and shows the same warning.
- Double-clicking a Workspace file opens it like a dropped file:
  .MID/.MIDI/.SCRAFT in any case open as MIDI/session (they used to
  import as audio), and a failed open is logged.
- convert OUT with an extension that is not valid Unicode is refused as
  an unsupported format instead of being written as WAV.

* fix(app): read Env after the launch flags; tidy Switch and Env

* fix(audio-io): decode every extension the format table lists

* fix(cli): write convert output atomically; tighten --connect and loopback

- convert writes every output to a new file in the output's folder, flushes it and renames it
  over the target, removing the new file on failure. A failed write (disk full, unwritable
  folder) no longer truncates the input under --in-place (RFC-0012 rule 1.6). A link output
  still writes the linked file.
- --save in place already goes through the engine's replace_file (tmp + rename), and
  file.bounce_mix renders to <name>.part and renames it when complete, so both are atomic.
- The desktop's --control / SOUNDCRAFT_CONTROL_PORT no longer treat [::1] as loopback: the
  control server binds 127.0.0.1 only. [::1]:P is a usage error from the flag and is ignored
  with a warning from the variable. The CLI's --connect / app --port still accept [::1]:P.
  Error tip and docs/cli.md updated.
- mcp --connect conflicts with --demo (exit 2): the demo session belongs to the in-process
  server. Help and docs/cli.md say so.

* test(audio-io): list the decode fixtures in ATTRIBUTION.md

cargo xtask assets needs a row for every committed audio file; one
directory row covers crates/audio-io/tests/fixtures/.

* fix(cli): refuse non-finite --start and --end values

* refactor(cli): parse the control target once; open sessions by path; keep error sources

- A1: ControlTarget { host, port } is parsed once by args::control_target; is_loopback() replaces
  control_port/ControlPort, and Remote::new takes &ControlTarget.
- A2: engine::io::open_session takes impl AsRef<Path>; the CLI no longer requires a Unicode path
  to open a session, only for the --bounce and --save paths that go into JSON.
- A3: ValueError::Json keeps the serde_json::Error; Remote::request returns RemoteError and
  AutomationRoots::new returns RootError, both carried by CliError.

* test(engine): pass session paths to open_session without a borrow

* test(cli): info opens a session at a non-Unicode path

* refactor(app): log the shared long version; info: session extension in any case; Switch beside Env
2026-10-27 21:09:43 -04:00

7.8 KiB

MCP server

SoundCraft speaks the Model Context Protocol over stdio, so Claude and other agents can make music with it. The server is the mcp subcommand of soundcraft-cli, which docs/cli.md describes.

soundcraft-cli mcp                    # headless engine (no window, no audio device)
soundcraft-cli mcp --demo             # headless, starting from the demo session
soundcraft-cli mcp --automation-read-root DIR --automation-write-root DIR   # headless, confined to folders
soundcraft --control 7801 &           # the desktop app…
soundcraft-cli mcp --connect 7801     # …driven over MCP: same tools plus screenshots and clicks

Claude Code:

claude mcp add soundcraft -- soundcraft-cli mcp --demo

Confining file access

soundcraft-cli mcp --automation-read-root /work/project --automation-write-root /work/project/out
soundcraft-cli mcp --automation-read-root /work/project --automation-read-root /work/shared-samples \
  --automation-write-root /work/project/out --automation-write-root /tmp/scratch
{"mcpServers": {"soundcraft": {"command": "soundcraft-cli",
  "args": ["mcp", "--automation-read-root", "/work/project", "--automation-write-root", "/work/project/out"]}}}

The two flags (VectorCraft and the other Craft apps take the same ones, so one client config works across the suite) keep the agent to the folders you give it:

  • Every file a command reads must lie inside a read root: open_session, session.open, import_audio and file.import_audio, file.import_midi, file.import_video, file.import_clip_groups, file.import_session_data, track.load_preset and mix.sampler_load. The media a session file names is read too, so an opened session whose audio paths lead outside the read roots loads without that audio and lists it under missing.
  • Every file a command writes must lie inside a write root: save_session, session.save, session.save_as, session.save_copy, session.save_template, bounce_mix and file.bounce_mix, file.bounce_stems, file.export_clips, file.export_midi, file.export_session_text, file.export_selected_tracks_as_session, file.export_sibelius, file.send_to_sibelius, file.print_score, file.export_clip_groups and track.save_preset, together with what a session save writes beside the session file (Audio Files/, Backups/) and its temporary file. A score command given no path writes to the temp folder, which is outside the write roots unless you made it one, so it is refused.
  • Each flag may be given more than once (#874): a path inside any of the read roots may be read, one inside any of the write roots written, so one session can read a project folder and a shared samples folder, or write to an output folder and a scratch folder, without widening a root to their common parent. A folder given twice counts once.
  • Read and write access are separate. Give the same folder to both flags for both; a kind left out grants none of its access (with only --automation-read-root, nothing can be written to disk). Without either flag nothing changes.
  • A path is made absolute against the server's working directory, then its links are followed: a file (or folder) link that leads out of a root is outside it. For a file that doesn't exist yet, its nearest existing folder's links are followed and the rest must be plain names, so .. below a folder that doesn't exist is refused, as is a link that leads nowhere (writing through it would create its target). The result is compared with the root by whole folder names: /work2/x is not inside /work. On Windows case is ignored, UNC paths (\\server\share\...) are resolved like drive-letter ones, and device names (CON, NUL, COM1...) are refused.
  • A refused path is an ordinary tool error the agent can read and correct, for example automation path rejected: outside the write root /work/project/out: /etc/x.wav (with several, outside the write roots /work/project/out, /tmp/scratch: ...) or automation filesystem access is not granted: write authority is absent: /work/project/out/x.wav. A refused save path is reported as refused, never as "already exists".
  • The roots must be existing folders; a missing one stops the server with an error. Each flag takes one folder per use (--flag <dir> or --flag=<dir>) and may be used again for more, and a misspelled --automation-... flag is an error rather than an unconfined server.

The flags confine the in-process server, so --connect together with them is an error: the server would otherwise talk to an app that isn't confined. SoundCraft's desktop app has no equivalent flags yet, so a running app (soundcraft --control) is not confined. --demo is the operator's own choice and is built before the roots apply.

Playback and bounce need no check of their own. SoundCraft decodes audio into the session when it enters it (import, open, sampler load), so rendering reads nothing from disk afterwards. Movies are the exception: a session only references them by path and the desktop app decodes pictures from that file, but the headless server never does, and the movie's path is checked when it is imported and when a session naming it is opened.

What the roots do not cover: commands that take no path still read and change the open session, and inspect_session returns it. Plug-in discovery scans the system's plug-in folders and loads plug-in binaries, which are fixed locations and native code, not agent paths. Recording is only reachable in the desktop app. Anything that reads or writes a path goes through one place in the engine (io::read_file, io::write_file and the helpers beside them in crates/engine/src/io.rs); a test keeps the rest of the engine from touching the file system on its own.

The checks hold for every thread of the server process. A path is checked, then opened: another program that swaps a link in between can still win that race, so the roots keep an agent to the files it was given rather than guarding against other software on the computer.

Tools

Tool Purpose
list_commands Discover every command (id, label, menu path, shortcut, params doc, enabled).
execute Run one command with JSON params.
batch Run up to 1000 commands in order (keep_going optional). Larger batches return isError: true before executing any command. Any failed call makes the tool result isError: true; with keep_going the remaining calls still run and every result is listed.
inspect_session Verify work: tracks, clips, mixer, plugins, sends, automation, markers, selection; each sources[] entry includes loaded (decoded audio in the pool).
new_session, open_session, save_session Session files (.scraft, audio in Audio Files/). open_session returns missing: the media files that could not be loaded (empty when all are found). save_session won't save over another existing file unless replace is true.
import_audio WAV, AIFF, FLAC, MP3, OGG, AAC/M4A, ALAC, CAF onto a new or existing track.
bounce_mix Render to WAV/AIFF/FLAC; returns peak dBFS and integrated LUFS.
parity Feature-parity report.
screenshot, ui_inspect, ui_set, menu_invoke, click, drag, key App only (with --connect).

Resources: soundcraft://session (full session JSON), soundcraft://commands.

Conventions

  • Programmatic calls never open dialogs; empty params act on the current selection.
  • Positions: samples, {"seconds": x}, or "bars_beats:5|1|000"-style strings.
  • Tracks by name or id; clips by id.
  • Every command is undoable through edit.undo unless it is a query or a view setting.

The acceptance test crates/automation/tests/it/agent_tasks.rs drives a realistic session end to end through MCP only (mixing, plugins, routing, editing with undo, markers, tempo, automation, bounce, save and reopen).