LangBeat

Use LangBeat from your assistant

LangBeat’s Model Context Protocol server connects ChatGPT, Claude, Codex and other compatible assistants to your account, library, playlists, music analysis, video projects and live Studio.

The connector uses your signed-in account. Source discovery and downloads follow your permissions. Automatic stem rendering is currently in private QC; get_create_capabilities tells your agent whether your account can render.

Connect with OAuth

Add https://langbeat.live/api/mcp to your assistant, then sign in to LangBeat and approve its permissions. Your assistant refreshes its access automatically. You do not need to renew the connection every hour.

Codex

Use this configuration, then run codex mcp login langbeat or choose Authenticate in Codex’s MCP settings.

[mcp_servers.langbeat]
url = "https://langbeat.live/api/mcp"
tool_timeout_sec = 120

# Sign in once: codex mcp login langbeat

Claude and other MCP clients

In Claude desktop or claude.ai, add the server URL as a custom connector. In Claude Code, run claude mcp add --transport http langbeat https://langbeat.live/api/mcp, then use /mcp to sign in. Other OAuth-compatible Streamable HTTP clients use the same server URL.

Replace an old hourly connection

Remove the old LangBeat Authorization header, bearer_token_env_var, or bearer-token setting before signing in with OAuth. Static credentials take precedence over OAuth in clients such as Codex. Replace the old configuration rather than appending a second LangBeat entry.

Choose permissions on the LangBeat sign-in screen. Manage or disconnect authorized apps in Create or Studio → Connect assistant. Active connections renew automatically; after 90 days without renewal, sign in again. A public catalogue API key is separate and does not authorize this connector.

Account and app access

The connector exposes 102 tools. Read access covers your profiles, preferences, music, playlists and app projects. Start with whoami to confirm the account and active profile.

To change profiles, settings, saved tracks, playlists, follows or video projects, reconnect and select Manage account and library. Existing read/edit connections do not gain this permission automatically. These actions apply directly; the Co-Producer mode (Listen, Riff or Lead) governs song edits only.

After an update, use Refresh tools in ChatGPT’s LangBeat settings. Password and email changes, billing, credential creation and account deletion use the signed-in app’s Account settings. The connector does not expose administrative maintenance operations.

Available tools

ToolWhat it does
get_create_capabilitiesCheck which song engines your account can use before starting work.
get_compositionRead the existing composition behind an owned private or readable public song: source bars, pitch/tempo transforms, channels, effect returns, mixer, automation, and optional transcribed notes/hits. Does not compose or publish. Private recipes use owner-scoped storage; missing sessions return explicit unavailability. Arrays paginate independently with eventOffset/eventLimit; use channelId to focus one channel.
find_stem_sourcesFind analyzed vocals or backing recordings, optionally matched to an anchor track.
get_track_stemsRead a recording's available separated stems, measured tempo and key.
get_track_genomeInspect a readable recording's named sounds and instrument timeline with confidence scores. Optional time window filters timeline spans, not the sound inventory. Paginate sounds and timeline independently; labels are hypotheses, not musical verdicts.
get_source_measurementsRead cached measured beat/bar timing and source-mix loudness, peak and clipping-heuristic evidence for an accessible recording. Missing measurements remain explicit. Does not measure per-stem Hz energy.
query_vocal_attributionRead existing source-bound singer/role research proposals for an accessible recording, optionally by time window. Returns documentary references, alternatives, timing provenance and audition links. Currently covers the five-recording QC study; does not analyze an arbitrary new song or recognize voices. All returned proposals are unverified and must not become published singer metadata without review.
measure_stem_energyStart actual PCM frequency-energy analysis of 1–8 accessible separated stems on an explicit timeline. Set source windows, destination offsets and gains; no automatic normalization or inferred alignment. Returns an account-bound receipt: poll get_stem_energy_job, then query_stem_energy. This starts compute; do not repeat while a receipt is pending.
get_stem_energy_jobRead the status and provenance of your remote stem measurement. A processing response is not a finished measurement.
query_arrangement_energyMeasure the actual rendered lanes of your completed authored song by time and Hz. Uses the saved render, including cuts, varispeed, automation, filters and derived returns. Energy is before constant master gain; returns are excluded from independent masking pairs. Exact boundary cropping defaults on. Only your own completed render job is readable; no file paths or new render required.
query_stem_energyQuery a completed measurement by time and Hz. Returns per-stem contribution, actual-sum interaction and timed possible-masking candidates. exactEdges recomputes partial bins from verified PCM. These are measurements, not musical approval. Receipt belongs to the signed-in account.
create_stem_songRender one selected vocal/backing pair or three fresh stem arrangements into your Private Tracks.
render_stem_arrangementRender your explicit per-stem arrangement into one private editable track. Declare accessible track/stem IDs, source cuts, destination times, pure varispeed rates, optional independent pitch shifts, fades, lane gain automation and optional derived delay returns. Every clip needs a musical reason. No automatic selection, alignment, normalization or musical approval. Reuse requestId after timeout, poll get_song_job, then verify list_private_downloads. Preview is an excerpt of the same full render, not a separate arrangement.
get_song_jobRead actual render progress and completed outputs. A queued job is not a finished song.
list_private_downloadsList your private Create songs and obtain short-lived listening download links.
find_live_sessionFind the song the person has open in LangBeat (Create or Studio) right now: their recent songs, which one has a live tab, and which view it shows (create or studio). Every live tool finds it on its own when compositionId is omitted; use this to see the choice or pick another song.
get_session_stateRead the open song as the person's tab has it: mode (Listen / Riff / Lead), view (create or studio), tempo, key, sections, the CHORDS row, lanes (name, kind, fragile, source record tempo and key, mixer, instrument patch and parameters), every region (source, window, tempo, transpose, key lock), the Sounds list, pending ideas and the change log revision. Read it before editing and after, to confirm what landed.
edit_sessionEdit the open song: up to 64 commands as one all-or-nothing envelope and one undo group, applied by the person's tab in their mode (Lead applies live, Riff makes one idea for them to accept, Listen ignores). At most 4 envelopes a second per agent: batch instead. The commands schema gives each command's fields, units and ranges. Regions: add_region, trim_clip, split_clip, duplicate_clip, set_clip (gain, fades, reverse), swap_source, move_clip, delete_clip. Lanes: set_lane, delete_lane, set_track_mix, set_track_effects, set_section_effects, set_lane_inserts (EQ, compressor, drive), set_sidechain, write_automation, write_notes, set_group. Song: set_master, set_sections, set_key, set_clock, apply_cast, set_bpm, set_title, new_composition. Pads: assign_pad, clear_pad, place_clip. Studio: set_instrument, write_drum_steps, set_part, set_chords. Measure a mix with render_check. Add a one-sentence `note` saying why. Returns applied, proposed, ignored, rejected, or timeout (unknown: read state before retrying).
undo_session_editUndo the last edits in the open song (Lead only; the person undoes their own in Riff and Listen).
check_vocal_cutsBefore cutting a vocal, check beats on a lane: each is snapped to the nearest breath (voiced and fricative bands dip together) with a verdict: clean, in_sound (move it), side_leak (a backing voice continues) or no_audio. Read-only; works in every mode.
plan_tempo_matchPlan one tempo and key for every lane (read-only): the fragile singer keeps time, other singers varispeed into a compatible key or stretch with their key kept, tonal parts stretch and shift, drums follow. Add `extra` records to see how they would fit. Apply with edit_session [{type: apply_cast}].
show_in_libraryShow the person what you are considering: switches their library panel to your shelf, query and record, and outlines your picks. Changes nothing in the song.
replay_sessionBuild an engine's song (compose_in_style, the recorded engine, or any published composition) in the person's workspace lane by lane: the song first (exact tempo, key, sections), then one envelope per channel, delayMs apart, so they watch each arrive or approve each in Riff. Long songs continue with fromEnvelope = the previous nextEnvelope.
search_librarySearch the person's saved library (all of it), with fit badges against the open song: the cast solve's move, rate, landing key and status (clean / listen / clash) for each record as a singer or an instrumental part.
find_phrasesBar-aligned windows of N bars where one stem of a record actually plays (from its loudness envelope and resolved beat grid). Each phrase carries a source ready for add_region or assign_pad, and per-bar or per-beat slices on request.
show_in_studioPoint the person at something without changing it or moving their view: a part (clipId), a lane, a bar range, or a Sounds row. Their tab offers it ("Claude wants to show you Lead, bars 13-16 · Show") and they choose to look. Works in every mode; never edits.
get_session_changesWhat happened in the open song since a revision: who (the person or which agent), which commands, which parts and lanes, the outcome (applied, proposed, declined, ignored, pending) and the note. Returns the newest revision; pass it next time. Read-only; it doubles as the audit log.
get_band_energyPer lane and per bar, low (below 120 Hz), mid (120 to 800 Hz) and high (above 800 Hz) energy of the song as it is now, rendered offline in the person's tab, plus the mix. Absolute dBFS, so bars and lanes compare directly: check that a chorus lifts before calling it done. Read-only.
render_checkAsk the person's tab to bounce the song (or bars) offline through the master and measure it: sample peak, true peak (dBTP), integrated loudness (lufs, ITU-R BS.1770 gated), RMS, clipped frames, stereo correlation and findings. Read-only; says unsupported when the tab cannot render.
import_soundAdd a stem from the person's library or a public LangBeat track to the song's Sounds list (external URLs are not accepted). Lead adds it; Riff shows it as proposed until the person keeps it; Listen ignores.
publish_sessionAsk to publish the open song (destination public or private). Never publishes on its own: the person's tab shows the Publish dialog filled in and only their click goes ahead. Lead only. Returns pending; their answer appears in get_session_changes.
genome_trackGenome one song on demand: cut and name its sounds (kick, snare, bass notes, loops) for the library's one-shots. Returns at once; follow with get_genome_status. A public song or one of the person's own; it needs separated stems.
get_genome_statusA song's Sound Genome status: ready (with its sound families), running (with the queue ahead), stalled, not_genomed or needs_stems.
keep_soundKeep a sound in one of the person's collections (Your Collection unless collectionId is given): a Song Genome sound by soundId (take original = the cut from the recording, rebuilt = its resynthesized twin), or a stretch of one stem by trackId + stem + startSec + endSec (cut on the server with 5 ms fades; snap the edges to the beat grid first, 120 s at most). The file is copied into the person's private space; keeping the same sound twice returns the one already kept.
list_collectionsThe person's sound collections (Your Collection first) and the public Starter collection, with counts, visibility and share codes; or, with collectionId, that collection's sounds (name, family, kind, source song and moment, tempo, bars, key, recipe, licence, playable url), up to 500 a page.
rename_soundName a Song Genome sound just for the person (name null clears it), or with forEveryone suggest it as a fix everyone sees once verified: at once if the person is an artist on the song, when three people back the same name and it is among the sound's own top labels, or by a LangBeat admin. Sounds are named by song, stem and moment (startSec), so names survive a re-analysis.
whoamiRead the signed-in LangBeat account, owned profiles and active identity. Never returns tokens or credentials.
list_profilesList the profiles owned by your signed-in account, including private profiles.
get_profileRead a profile visible to your account. Private profiles belonging to other accounts are unavailable.
update_profileUpdate the name, handle, biography or visibility of one of your profiles. Only supplied fields change. Public profile edits are immediately visible.
create_profileCreate an additional listener or creator profile on your account. Starts private; choose public visibility later with update_profile. Does not grant verification or paid entitlements.
delete_profilePermanently remove an inactive, non-primary profile you own. Cannot delete the primary, active or last profile. Requires explicit confirmation.
switch_active_profileSwitch the account’s active profile. Subsequent saves, playlists and social actions use this identity; the app reads the same setting.
check_profile_handleCheck a handle against the app’s shared profile and artist namespace before changing it.
get_profile_statsRead the app’s listening, library and creator statistics for one of your profiles.
get_account_securityRead your sign-in providers and email-verification status. Password, email, billing and account deletion remain in LangBeat’s signed-in Account settings; no passwords or keys are returned.
get_preferencesRead stored playback preferences: theme, transitions, shuffle, repeat and song sharing. Null means the app uses its defaults.
set_preferencesUpdate only the supplied playback preferences. Transition choices are marked as a deliberate user choice.
get_user_settingsRead cross-device player appearance, mixer layout, deck count and blend preset.
set_user_settingsUpdate supplied cross-device player settings while preserving other stored settings.
search_profilesSearch public profile handles and display names. Private profiles are excluded.
list_profile_followsRead profiles followed by your active identity, or its followers. Profile identities remain separate.
follow_profileFollow (vibe with) a public profile as your active profile. This changes your social graph.
unfollow_profileRemove a follow (vibe) from your active profile.
list_connected_appsList OAuth assistants connected to your account and their granted permissions. Does not expose tokens.
revoke_app_connectionDisconnect one OAuth assistant and revoke its session. Disconnecting this assistant will stop subsequent calls. Requires confirmation.
list_developer_keysList your developer API key names, prefixes and usage. Secret key material is never returned.
search_tracksSearch readable tracks by title/artist and tempo, key or genre. Public and owned tracks only; mine restricts to your uploads. Use returned IDs for saves and playlists.
get_trackRead metadata and processing status for a public track or one of your own recordings.
list_my_tracksList recordings owned by your account, including private songs. Audio download links for private Create songs are available through list_private_downloads.
update_trackUpdate the title or genre of a track you own. Does not change ownership, rights, artist credits, audio or publication status.
list_saved_tracksPage through your saved library, newest saves first. Playlist membership is separate from saved-library membership.
save_tracksSave readable tracks to your library under the active profile. Already-saved tracks are left unchanged.
unsave_tracksRemove tracks from your saved library. Does not delete recordings or change playlists.
get_listening_statsSummarize your actual listening events over a bounded time window. Reports when the 5,000-event sample is truncated; never invents missing history.
list_playlistsList your playlists with their visibility, pin state and profile identity.
get_playlistRead one of your playlists and a page of its tracks in playlist order.
create_playlistCreate a private playlist under your active profile. Reuse playlistId on retries to avoid duplicates.
update_playlistRename, pin, or change the visibility of a playlist you own. Making it public exposes its track listing.
delete_playlistPermanently delete one of your playlists and its membership list. Recordings remain intact. Requires explicit confirmation.
add_tracks_to_playlistAppend readable tracks to your playlist atomically, preserving existing order and skipping duplicates.
remove_tracks_from_playlistRemove selected tracks from your playlist. Does not delete recordings or remove library saves.
reorder_playlistAtomically reorder your playlist. Supply every current track ID exactly once; a stale or incomplete list is rejected without changing order.
sort_playlist_for_mixingSuggest a tempo and Camelot-key running order for your playlist. Set apply=true to persist atomically. Unknown tempos are reported. Supports up to 500 tracks.
get_share_linksGet app links for a readable track or your playlist. Does not publish anything; reports whether the item is public or marked shareable.
get_track_lyricsRead a track’s existing lyrics, timing and source/alignment provenance. Missing or unaligned lyrics remain explicit.
sync_track_lyricsFetch or refresh lyrics for a recording you own through the app’s existing lyrics pipeline. This is not manual transcription or proof of alignment.
separate_stemsStart stem separation for a recording you own, subject to the app’s quota and admission checks. Poll get_stem_job; a queued response is not completion.
get_stem_jobRead stem-separation status for an accessible recording.
analyze_track_structureRequest the app’s structure analysis for your recording. Existing jobs and quota limits are respected. Read get_track_structure for measured results.
get_track_structureRead measured sections, beat timing and structure-analysis version for an accessible recording. Null means no measured result yet.
import_track_urlImport a supported YouTube, Spotify or Apple Music link through LangBeat’s existing import pipeline. Uses your quota. Poll get_import_job instead of repeatedly submitting.
get_import_jobRead status and the resulting track ID of an import job owned by your account.
list_import_jobsList your recent music imports and their actual processing outcomes.
list_compositionsList saved compositions owned by your account, including songs without a live Studio tab.
get_create_preferencesList available source artists and genres for the app’s song-generation preferences.
search_artistsSearch public artist pages by name. Returns artist IDs and slugs.
get_artistRead an artist page’s public identity and biography by slug.
get_chartsRead the app’s current regional music chart. Results retain their source and ranking provenance.
get_trendingRead LangBeat’s trending tracks for a time window and optional city.
get_your_setRead the personalized set the app builds for your active profile from its saved library, listening history and region.
list_video_projectsList video-edit projects owned by your account.
get_video_projectRead one of your video-edit projects, including saved clips, effects and planner parameters.
create_video_projectCreate a private video-edit project using the same validation as the app. Returns its project ID; do not retry after an unknown outcome without listing projects first.
update_video_projectUpdate supplied fields of your video-edit project. Clips and effects replace their corresponding saved arrays when supplied. The app validates the edit.
list_video_assetsList your uploaded video assets and short-lived preview URLs, optionally restricted to one project. Does not import footage.
get_video_renderRead the actual status and output of your video render, or the latest render for your project.
start_video_renderSubmit an explicit EditPlan for your video project to the app’s cloud renderer. The app validates the plan, asset ownership and quota. Poll get_video_render; submitting does not mean completion.
reel_photo_uploadsStep 1 of a Reel from photos: one signed upload per photo, in any order (the Reel picks the best order itself). PUT each photo’s bytes to its uploadUrl (JPEG, PNG, HEIC or WebP), then call make_reel with the batch.
make_reelStep 2: cut an Instagram Reel from the uploaded photos and a song in your library ("Title - Artist" or a track ID), on the song’s beats around its chorus. Auto is the neon edit with a new look each Reel; look pins parts of it; style picks an older template instead. Uses one video render. Poll get_reel; queued is not done.
get_reelStep 3: a Reel’s status and, once complete, its video and contact-sheet links and the check results (READY when every cut is on the beat and the audio is in sync). Without renderId: your recent Reels.
get_playlist_collaboratorsRead the collaborator roster and pending requests for an accessible playlist through the app’s permission checks.

Inspect the music before arranging

Use get_track_genome to inspect named sounds, confidence scores and instrument changes in a passage. For example, set startSeconds: 88 and endSeconds: 104 to inspect 1:28–1:44. Follow nextOffset and nextTimelineOffset independently to retrieve the full sound inventory and timeline. The time window filters timeline spans; sounds remain a paginated inventory.

Use get_source_measurements for cached beat and bar timing, source loudness and peak measurements. Missing measurements are reported explicitly. Genome labels are model estimates and these measurements do not establish musical compatibility.

For fresh frequency measurements, call measure_stem_energy with 1–8 stems, their track IDs, stem types, source windows, timeline offsets and gains. Poll get_stem_energy_job with the returned receipt. Once complete, query_stem_energy measures any time and frequency range, including relative contributions, summed-audio interaction and candidate masking intervals. Set exactEdges: true to recompute partial boundary bins. Receipts are account-bound and expire after 24 hours; your OAuth connection renews independently.

Try a three-song session

Check LangBeat’s Create capabilities. Make three fresh songs using separated stems, with different source recordings. Keep them private. Wait for completed renders, then show each song’s title, duration, tempo, sources and private download.

The agent checks availability, inspects sources, and starts create_stem_song with count: 3 and stable request and project UUIDs. It polls get_song_job every 15–30 seconds, then verifies all three outputs through list_private_downloads. A queued job or a partial result is not three completed songs.

Choose your own vocal and backing

Use find_stem_sources with the vocal role first, then search the backing role with that recording’s anchorId. Inspect the separated parts using get_track_stems. Pass the two recording IDs to create_stem_song with count: 1. Compatibility uses recorded analysis; a matching BPM or key alone does not prove that a pair sounds good.

Private downloads and ownership

Finished songs appear in Create → Your Songs, where you can search, listen and download. Download links expire after ten minutes. App management tools can update your profiles, settings and playlists. Deleting a playlist or inactive profile requires explicit confirmation; recorded songs are not deleted or published by these tools. Source recording rights still apply; stem separation and private creation do not grant a licence to release a remix.

How this differs from the catalogue API

The LangBeat REST API searches public music measurements and charts. The MCP connector performs authenticated account, library, analysis and creation workflows. For music concepts, see stem energy, harmonic mixing and measured BPM and key.