Live Smith

An AI production assistant that works with your Ableton Live.

What it does · Getting started · Connections · Safety · Privacy

Live Smith helps you understand, arrange, and edit a Live Set through conversation. Select a track, Clip, device, or another supported Live object and choose Ask Live Smith. Your conversation starts with that musical context.

The compact Context row inside the composer shows the bound object or opening selection. Select it to open the detailed Context view. Positions use Arrangement beats when available; the strip does not follow later Live selections or restrict edits to the displayed object. An unavailable object is labelled explicitly. Use the Session's Scope control to choose allowed edit categories.

Describe what you want to do in the message box and continue the same conversation as the work develops. The context strip and detailed Context view remain available without changing the draft or choosing a task mode. The message box starts at one line, grows with its content, and scrolls within a bounded height for long drafts so conversation history remains visible.

The Inspector separates Session views (Context, Artifacts, Skills, Tools) from global Settings (Agent, Extensions, App). Context contains the creative brief, bound Live object, context usage, and the next message's attachments. Skills selects workflow instructions for the current Session. Tools automatically loads its directory when the dialog opens and refreshes it when the Session or tool sources change. Discovery waits for foreground work to finish and leaves the composer usable. Reload tools refreshes descriptions; a failed load can be retried. Only enabled, approved MCP sources are contacted, and no tool is executed. The directory is a snapshot; the next request's inputs and connection settings can change which tools are available. Expand an MCP tool to edit its supported parameters and select Run tool. Cached tool parameter forms remain editable during generation; Run tool becomes available when the current operation finishes. Native controls use the tool's defaults, choices, and value limits. The tool runs directly, and its result appears in the panel and Session history without a model request. Use the composer's Stop control to cancel a running tool. Parameter drafts last within the current dialog and Session; switching Sessions or changing connections resets them. Tools that provide an MCP App also offer Open interface. The Plugin supplies the layout and interaction; its interface runs in an isolated frame and uses the same approved MCP server. Closing the interface cancels pending operations. Completed tool results offer Use in chat, which adds a reference to the composer. Saved MIDI offers Add to Live, opening a separate import dialog that reads the file's source parts and current Live tracks automatically. Select the parts to import, use new MIDI tracks or map them to existing destinations, and choose a common Arrangement start beat. A single part can use the Session's bound MIDI track; a bound Arrangement range supplies the suggested start. Merge all parts into one Clip combines the complete file explicitly. The preview shows new tracks and resulting Clip boundaries before the ordinary Session approval flow. New tracks require the corresponding Edit Scope. Tempo, meter and controller events remain in the saved file; import writes notes only and leaves the Set's tempo and meter unchanged. An interrupted import preserves recovery information and must be inspected before retrying.

Session → Artifacts collects this Session's saved MIDI, downloaded service audio and Plugin WAV/MP3 results. Other supported file formats remain chat attachments. Each MIDI or audio work has one entry containing its versions. Open the entry and choose a version to inspect, export, attach or add to Live. The version's source is shown explicitly; the source can be an older version in the same group. Make primary saves the work's default version; reopening its entry selects that version, while an explicitly selected version remains the target of export, attachment and further work. Clearing the primary returns the default to the latest available version.

For MIDI, Version comparison defaults to the source version and allows any other version in the same work as the comparison baseline. It reads both complete MIDI files and shows a short change summary and an overlaid piano roll for the selected part. Uniform transposition is identified across complete parts; other summaries cover pitch, timing, length and velocity changes. The chart marks the two versions on one beat/pitch scale. Newly saved model and Plugin MIDI results also appear in chat inside the existing tool result. Expand it for a compact note preview, part selection, Export MIDI, Add to Live, and Open artifact actions. Each card refers to its exact saved version. Saving MIDI does not edit Live or require Live edit approval. Older text-only results remain available through Artifacts. Both MIDI preview and comparison use the same scrollable piano roll: zoom in or out, move horizontally, focus the first notes/changes, or choose Full view. Long files initially show a 16-beat window near the first relevant note. The compared files retain their actual timing and length. Renamed parts can retain their correspondence through a unique MIDI channel; ambiguous parts or notes may remain additions and removals. Tempo, meter and controller events are not compared.

Built-in generation alternatives share a work when their output kinds are compatible. Revisions retain their explicit source across resume and later download. Stem files remain separate components with source provenance; they are not numbered as alternative versions of a complete song. MIDI and audio files have separate version groups even when one was produced from the other.

Preview part isolates one MIDI track/channel part. Opening an artifact loads its complete notes within the saved-file limit; catalog entries retain a small overview. It does not alter the saved file or export; Add to Live carries the part choice into the import dialog.

Attach to message and Export are available for both MIDI and audio. They reuse the exact saved file without calling a model or audio provider. Repeated attachment of the same pending file reuses its existing reference; distinct MIDI versions and audio outputs remain separate. Model input support and attachment budgets still apply when sending. A missing historical tool call leaves its generation parameters unavailable. Refreshing the list retains the selected version and audio playback; Back to chat pauses artifact audio and returns to the composer.

Continue in chat (or Create next version for MIDI) selects a source for the next chat request and adds a draft to the composer. Add the desired changes and send it through the usual model and tool permissions. The source is consumed when that initial user message is saved; Steer and manual tool calls do not consume it. The resulting tool calls retain their parent artifact even if a later request uses another source. Clearing the next-request source leaves the original files intact. MIDI artifacts use the same mapped import preview; audio import is prepared in chat and uses the existing scoped audio action and approval preview. The artifact library does not audition Live instruments or roll back applied changes.

Ask the current model to save a MIDI artifact to generate a file without changing Live; no external MIDI generator is required. MIDI revisions are grouped as v1, v2, …. Creating a revision preserves every previous file, including when starting from an older version. Attach to message adds the selected version to the next message as a MIDI attachment. Export MIDI downloads that exact version as a standard .mid file with its original tracks, channels and events. Import the file into Live, Cubase or another compatible DAW, or drag it from the file manager. Direct file dragging out of the extension panel is not supported. Export does not include instruments, effects or rendered audio.

Session → Tools → MIDI continuation builds an ordered buffer of up to four future sections from selected MIDI Clips. Choose the current Session model or an approved local Plugin/MCP tool that declares MIDI conditioning, then save the source selection, section length and capacity. Fill buffer generates only the vacant slots; each section uses the original context and its preceding section. The current model uses the Session's saved creative brief. Import next section offers the usual part-to-track import preview. Only a successful import advances the buffer; cancelling the preview leaves it intact. Fill buffer continues from the last saved section, and Stop retains completed artifacts.

Source notes, Clip markers/loops and Live tempo are checked again before generation and import. Changed sources or generator configuration require a new setup. The buffer has no playback clock or automatic Clip launch: auditioning and advancing sections are explicit actions. Source capture exports nominal MIDI notes with marker cropping and bounded loop expansion; instrument sound, probability and velocity randomization are not rendered. The context is limited to 16 source Clips and 4096 expanded notes. Each future section starts its own version group; revising one of its saved artifacts does not replace a queued section. Local tool conditioning retains original and previous parts as separate SMF tracks within the existing 32-track/4096-note file limits. Plugin output must fit the 960-PPQ conditioning format. Same-pitch overlaps and notes that collapse to zero ticks are rejected before entering the buffer.

Settings → App → Chat shortcuts controls the entries above the message input. Context, Brief and Artifacts appear together by default; Skills and Tools can be enabled individually. Brief opens the creative brief section in Context; other shortcuts open their Inspector panels. Hiding shortcuts preserves all Inspector tabs, saved content, enabled Skills and tool permissions. The preference is shared across windows.

Choose Settings → App → Interface language to use English, 简体中文, or follow the system language. The preference is shared across Live Smith windows. Switching languages keeps drafts and ongoing work; user messages, model replies, object names, and raw provider/SDK output stay in their original language.

Note

Live Smith is beta software. See the development guide for supported Live versions and installation from source.

What it does

• Work with your Set. Inspect tracks, Clips, devices, and MIDI notes, then create or edit MIDI parts, organize tracks and Scenes, adjust mixer settings, and work with observed samples and native Live devices. Analyze isolated Arrangement audio ranges for level measurements. With a verified audio-input model, Live Smith can also render a requested Arrangement Clip range directly into the active request for listening or transcription. Return and Main tracks can be inspected and used for bounded device-chain and mixer-parameter edits. Existing Rack Chains, including empty Chains, expose their direct devices and Volume, Panning, and Sends; ordinary Racks can append new empty Chains.
• Keep conversations organized. Use separate Sessions for different parts of your Set, revisit their history, or collapse the Sessions panel for more space.
• Choose models in the composer. Save several models in one connection Profile, then switch models and supported reasoning levels below the message box. The context meter shows the active model context when reported. Each model can override its context-window size and auto-compaction threshold; a blank threshold defaults to 90% when the context window is known.
• Guide the musical approach. Enable arrangement Skills per Session or mention one for a single request. Open a built-in Skill to read its instructions before enabling it in Session → Skills. Settings → Agent → Custom Instructions stores standing creative, workflow, and tool preferences across Sessions; the current request can choose a different workflow at any time.
• Install extension packages. Settings → Extensions → Plugins accepts portable Agent Plugin ZIPs and compatible Codex or Claude Code packages. Installation runs nothing: enable the Plugin, review each MCP server, and grant artifact input or output separately. Platform-only commands, hooks, agents, and apps are reported as unsupported rather than executed.
• Connect MCP tools directly. Settings → Extensions → MCP accepts a Streamable HTTP endpoint or a local stdio command without a Plugin package. Review the endpoint or process before enabling it; loading the Tools directory discovers its tools without invoking them.
• Bring reference material. Paste or drag images, documents, or audio into the composer. Input support depends on the model and connection.
• Keep work moving. Queue a follow-up for the next turn, steer the response already in progress, or stop it. Long runs of tool and Apply activity collapse into one expandable timeline item so the conversation stays readable. When a provider returns a visible reasoning stage or summary, Live Smith shows it in a separate Thinking item. Visible content is expandable; a stage with no text shows Thinking while in progress and disappears when complete. See visible reasoning output for connection-specific summary behavior.
• Search when needed. Compatible Direct API connections can enable hosted Web Search, with search activity and citations visible in the conversation.
• Generate music and sound effects. Add an ElevenLabs connection under Settings → Extensions → Audio services, then describe the music or sound you want. Multiple named audio connections, including separate accounts at the same provider, can be enabled together.
• Run audio tools from a parameter panel. Open Session → Tools, choose a named audio connection, expand a tool and select Open controls. The form offers that connection's supported inputs, including custom lyrics, sliders, sound parameters and source selection. Run tool invokes the audio service directly without a chat-model request. Results and tool history belong to the active Session; generation can consume the connection's allowance. Form drafts last while that Session, account and connection configuration remain active. Saved connection settings continue to supply credentials and the default model.
• Generate songs or instrumentals with Mureka. Add a Mureka connection with an API key from Mureka API Platform, then use prompt-based music generation. Live Smith polls the accepted task, saves the returned track locally, and can resume a missing download without submitting the generation again.
• Generate music with Google Lyria. Add a Google Lyria (Gemini API) connection with a Gemini API key. Choose lyria-3.5 for full songs, lyria-3-clip-preview for fixed 30-second previews, or the experimental lyria-realtime-exp transport for bounded instrumental generation. Realtime PCM is saved as an ordinary WAV result, so preview, model listening, and Live import use the same Session audio workflow as other generators.
• Generate through the official Suno Platform API. Add a separate Suno Platform (official API) connection and use an API key managed at platform.suno.com. Platform access and usage are separate from the consumer website subscription.
• Use an optional third-party Suno service. The explicitly labeled SunoAPI.org connection uses that provider's own key, billing, and callback. It is neither Suno Platform nor Suno subscription access.
• Connect a Suno.com account. Open Suno in your system's default browser, then explicitly import its Suno request Cookie in the local connection form. Both __client and current __session + __client_uat sessions are supported. Live Smith verifies, privately saves, and refreshes only the required fields without a browser extension or access to your Google credentials. Enable the saved experimental connection to generate music using your subscription credits, supply lyrics/styles and supported sliders, select male or female vocals, request a supported 10–480 second duration, browse songs/models, extend songs and get whole songs. The same connection offers Sounds with loop/BPM/key controls, Cover, Remaster, audio upload, Add Vocals, Add Instrumental, section replacement, native stem extraction and lyric writing. Source permissions and model capabilities are checked before submission. Replacement candidates are finalized through a separate explicit operation. Load account versions in the connection editor and choose a fixed version or follow the account default. The active Session's generated and processed audio appears above the chat composer, separately from connection settings. Completed songs generated through Live Smith can be auditioned through Suno's embedded player without downloading them. The embedded player is for human listening and is never model input. Download to Live Smith saves the selected song in its Session, after a separate confirmation that may consume one download allowance; Export MP3/WAV saves another copy through the default browser without requesting the audio service. Live Smith never buys extra quota. A verified audio-input model can then listen to that saved Session asset when asked. On macOS 14 or later, a generation challenge opens Live Smith's own verification window. Complete any challenge yourself; a successful result continues the original request once. Cancel or close the window to stop before submission. Initial sign-in still happens in your normal browser. Other hosts can generate only when Suno does not require a challenge. This is not full Suno website parity; see the supported features and limits in the website sign-in workflow. The connection uses an unofficial website protocol and remains subject to Suno's current terms.
• Separate audio into stems. Add a LALAL.AI connection, then ask to extract vocals, drums, bass, piano, or guitars from an audio attachment or an isolated Arrangement Audio Clip range. Saved results can be previewed, reused in later requests, and imported through ordinary Live edits. A verified audio-input model can also listen to a chosen saved stem when asked. Separation itself needs tool support but does not require model audio input.

Try requests such as:

“Create a four-bar bass MIDI idea on this track.”

“Help the chorus stand out using the parts already in this Set.”

“Inspect this device and explain what its current settings are doing.”

“Separate this Clip into vocals and drums, then place the results on new audio tracks.”

“Use my ElevenLabs connection to generate ten seconds of instrumental ambient piano.”

“Use my Mureka connection to generate an instrumental synthwave idea.”

“Use my Google Lyria connection to generate a 30-second instrumental preview.”

Audio processing is disabled by default. Each connection has its own saved key and uses that provider's allowance, separately from chat-model usage. Generation sends the requested description to the selected service; separation uploads the chosen audio and consumes LALAL.AI processing minutes for each requested stem, also returning the residual mix. Saved results can be previewed and imported through ordinary scoped Live edits. See external audio tools for supported inputs, limits, and recovery.

Composer commands are recognized only at the start of a message:

Press Enter to submit the composer. Shift+Enter inserts a line break, and the existing Cmd/Ctrl+Enter shortcut remains available.

• /compact [instructions] compacts the current Session now and can name what the checkpoint should preserve. The Session must have no active request and new conversation activity since its latest checkpoint. While manual compaction is running, the composer’s Stop control requests cancellation; once checkpoint persistence has begun, a successfully saved checkpoint takes priority.
• /steer <message> guides an active response immediately; while idle, it starts an ordinary request.
• /queue <message> schedules a turn after the active response; while idle, it starts an ordinary request.
• /clear switches to a fresh Session. The previous Session and any work still running there remain available in History.

Getting started

1. Install and run the extension using the development guide.
2. In Live, right-click a supported object and choose Ask Live Smith.
3. Open Settings → Agent, create a named Profile, and choose a connection.
4. Use Load Models, choose a default model, and Save & Use the Profile. Direct API connections also allow entering a model ID manually.
5. Ask for help. Review proposed edits according to the Session’s approval mode.

A Profile groups one connection and its model settings. Once it is saved, the composer lets you change the active Session’s model without returning to Inspector.

See model settings for catalog loading, generation options, and capability indicators.

Model connections

Direct API

Connect an OpenAI, Anthropic, or compatible API service using its endpoint and API key. Supported request formats are OpenAI Responses, OpenAI Chat Completions, and Anthropic Messages. API usage is billed by the provider.

Google Gemini works through its official OpenAI Chat Completions compatibility endpoint. See the Gemini setup for the connection fields and capability settings.

Hosted Web Search is available through supported OpenAI Responses and Anthropic Messages connections. It is off by default and configured per model.

Account subscription — experimental

Sign in inside Live Smith with ChatGPT, Claude, or Google Antigravity. Live Smith owns the OAuth session and calls the provider product backend directly; customers do not install Codex CLI, Claude Code, Gemini CLI, Antigravity, or provide an API key. Each subscription Profile has its own sign-in state, even when another Profile uses the same provider. Creating a new Profile therefore starts signed out, and signing out affects only that Profile.

After the provider returns a pending authorization, the Extension Host opens the default system browser on macOS or Windows. ChatGPT displays its device code in the dialog. Claude completes browser PKCE through a local callback. Antigravity returns to Google's hosted callback page; copy the authorization code shown there into Live Smith to finish sign-in. Live Smith checks the account automatically after that submission. If Google requires an additional account verification before enabling Antigravity, the dialog shows the verified Google page and asks you to sign in again after completing it.

ChatGPT uses the Codex backend API, Claude uses OAuth-authenticated Anthropic Messages, and Google uses the Antigravity product backend. Anthropic currently assigns third-party OAuth traffic to Claude Extra Usage when it is enabled. Antigravity uses the account's default entitlement and region; Live Smith does not import CLI-local license-tier or project-region overrides. Hosted Web Search is not exposed through subscription Profiles. If an account check is unavailable, use Sign out to clear its saved OAuth session before signing in again.

See model connections for setup requirements and provider-specific limitations.

Network proxy

Settings → App → Network Proxy provides three global modes: No proxy, System proxy, and Manual proxy. The selected route applies consistently to Direct API requests, subscription sign-in, token refresh, model catalog and model traffic, and external audio-service HTTP or WebSocket traffic. Loopback endpoints remain direct so local model servers keep working.

System proxy discovery follows static macOS HTTP, HTTPS, or SOCKS settings and the current Windows user's static Internet Settings. It uses a fixed, read-only Windows registry query; it does not run a shell or modify the registry. Automatic PAC/WPAD configuration is not evaluated. An automatic setting found in this static key is rejected; connection-specific or machine-scoped settings are not read and provide no static route. PAC/WPAD URLs cannot be entered in Manual mode; enter the concrete proxy URL instead. Windows System mode supports static HTTP and HTTPS destination entries through ordinary HTTP proxies; use Manual mode for an HTTPS proxy transport or SOCKS5. Manual proxy accepts an http://, https://, socks://, or socks5:// URL without embedded credentials. Existing installs remain on No proxy until this setting is changed.

Proxy modes are strict: Live Smith does not silently fall back to a different route. If the selected Manual or System proxy cannot be reached, provider requests stop with an actionable proxy message so you can start the proxy, correct the setting, or choose another mode.

You control the changes

Each Session has separate Approval and Edit Scope controls in the composer, initially labelled Manual and All scopes. Changes save immediately. The ? hints inside Scope explain the less obvious category boundaries.

Edit Scope	Allowed changes
MIDI	MIDI Clips, notes, and Clip properties.
Audio	Audio Clips, Warp, and Clip properties.
Devices	Instruments, effects, Racks, Drum Pads, and Simpler samples.
Mixer	Mixer parameters, mute, solo, and arm.
Structure	Tracks, Scenes, Cue Points, Take Lanes, and tempo.

Scopes can be combined. The All checkbox selects or clears every category; clearing all or choosing Read only inside Scope makes the Session read-only while keeping inspection available. To allow edits again, select the categories you want to permit. New and historical Sessions without a saved scope selection use All. The scope controls edits, not which Live information the assistant can read. Instruments and effects share Devices because the current SDK cannot reliably classify every device.

Live Smith can write MIDI or audio Clips into an existing Take Lane after it has inspected that lane. Take Lane Clip content uses the MIDI or Audio scope; creating or renaming the lane itself uses Structure. A newly created lane must be inspected before a later request writes into it, and Live Smith refuses to create over an occupied range whose overlap behavior cannot be verified.

Rack Chain creation uses the Devices and Mixer scopes because each new Chain owns both a device container and a Chain mixer. Existing Chain mixer parameter edits use Mixer. Drum Rack pad creation remains the dedicated Drum Pad workflow, which verifies the receiving note and reports partial completion. The current SDK does not expose Chain names, deletion, duplication, or reordering.

Plans outside the selected scope are rejected before any action runs. Container operations also need permissions for their contents: deleting or duplicating a track requires structure and mixer permissions, plus the scopes of its Clips and devices. Approval cannot grant a missing scope.

Within the selected scope, the approval mode controls confirmation:

• Manual asks before every proposed edit plan.
• Low Risk applies lower-risk plans automatically and asks before protected actions such as deletes, Clip writes, and sample replacements.
• Accept Everything automatically approves all authorized, validated plans, including deletes and replacement writes within the selected scope.

All modes still inspect the relevant Live state, validate actions, and check that the Set has not changed before applying an edit. One approved plan may create more than one Live Undo step.

Live edits use the existing expandable tool activity and retain their proposal and execution result in Session history. A single operation has one disclosure; consecutive operations share a group. Manual shows Apply and Cancel; automatic approval runs without those buttons. Applied, cancelled, failed, and partially applied operations have distinct states. Approval alone never counts as a completed edit.

For a single supported MIDI Clip creation, note edit, or device/mixer parameter edit, expand the card to view the observed Before and Proposed after. MIDI previews use Clip-relative beats and the same pitch/time scales on both sides. Large previews show at most 256 notes per side and state how many were omitted. Parameter previews show raw SDK values and observed ranges; they do not guess display units or map value labels to numbers. The full action list remains visible. MIDI creation previews cover empty Arrangement or Session destinations and exact reusable MIDI Clips. Plans involving several actions, other new objects, overlapping Arrangement Clips, or unavailable before/after data use the action list without a preview. A preview describes the proposal, not a completed edit or an audio audition; the selected approval mode still applies.

Scope changes are saved per Session and synchronized across open dialogs. You can change them during a request; queued plans and subsequent actions recheck the saved permissions. An action already in progress may finish, and completed changes are not rolled back when permissions are narrowed.

Live Smith does not run arbitrary model-generated code. It does not inspect or edit Automation, browse installed presets, or load a VST by plug-in identifier. Existing devices can be inspected and edited where Live exposes their parameters. Return and Main tracks intentionally exclude Clip, Take Lane, Arm, mute/solo, rename, duplicate, and delete-track actions.

Sessions, Skills, and attachments

Sessions keep conversation and action history with their Live context. Opening the dialog or choosing New Session does not save an untouched empty conversation. Messages, Session settings, and attachments are saved when used; the Sessions list keeps empty entries that were active in the current window and hides unvisited empty entries across tracks and History. Closing the window clears that temporary visibility. Conversations and unsent drafts remain visible; hiding empty entries does not delete existing data. Previous Sessions can be restored explicitly; matching names alone do not make an old conversation the same Live object.

Creative brief keeps the current Session's style, references, section structure, track roles, and material to preserve in one editable document. Open Session → Context, or use the Brief shortcut above the message input, then choose Save brief. The limit is 8,000 characters. Saved briefs remain available after model changes and context compaction. The model can offer a suggestion; Edit suggestion puts it into a local draft and Save brief explicitly accepts it. Suggestions never save preferences automatically. Unsaved drafts remain separate for each Session in that window. If another window changes the saved brief, review its current text before keeping or replacing your draft. The brief records creative intent; current BPM, meter, and other Live facts are read from the Set.

Hover over a message, or focus its controls with the keyboard, to copy its original text. Use as draft puts a user message back in the composer; it does not change history or undo Live edits. Choose Send after editing. Paused failed requests also offer Edit and resend. Editing their restored draft replaces that pending request instead of sending its old text again. Existing drafts require confirmation before replacement; historical attachments are not reattached automatically. Text drafts remain editable during generation and attachment upload.

Skills provide musical workflow guidance. Three built-ins cover section energy, musical variation, and instrument roles. They start disabled; View opens the full instructions without enabling them. Import and manage standalone SKILL.md files in Settings → Extensions → Skills. The library groups built-in, user, and Plugin-provided Skills by source, including disabled Plugin packages. Enable up to four Skills in Session → Skills, or use $skill-id for one request. Skills do not grant additional permissions or tools.

Plugins are installed packages that can contribute namespaced Skills and MCP tools. Manage packages in Settings → Extensions → Plugins; each package links to its capabilities in MCP and Skills. A Plugin starts disabled. Local MCP servers require explicit approval and run as your operating-system user; Live Smith does not provide an OS sandbox. Remote MCP servers also require approval. A Plugin tool cannot edit Live directly: declared MIDI/audio inputs are staged as temporary read-only files, and declared MIDI or WAV/MP3 outputs are validated and saved to the Session. Import remains a separate scoped and approved Live action. Removing a Plugin does not remove already saved Session artifacts. Remove a Plugin's connections and stop referencing its Skills before deleting the Plugin.

Plugins can declare persistent Plugin parameters on their package card. Save commits values for subsequent requests; sensitive fields remain write-only. Parameters survive restarts and package updates, and selected Skills and MCP configuration can reference them through ${user_config.KEY}. Discovered MCP Apps can also be opened from the Plugin card. See the configuration and UI author guide for supported formats, scope, and a runnable offline example.

Connections save named accounts or servers. Settings → Extensions → Audio services manages audio accounts; MCP manages directly configured servers and servers provided by Plugins, including their permissions. One source can have multiple named connections; a Skill-only Plugin needs none. Package MCP servers without credential fields can run anonymously under their server approval. Adding a named Connection for that package version and server selects its account configuration. Select Edit to configure a connection. Its bottom action bar places Remove on the left and Discard / Save on the right. Removal requires confirmation; unchanged saved connections disable Discard and Save, and new connections have no Remove action. Switching between Audio services and MCP preserves nonsecret drafts and clears newly entered credentials. In Extensions → MCP, Add connection opens the direct server configuration. For a Plugin server, use Add connection beside that server. Direct connections support Streamable HTTP with a URL, or a local process with a command and individual arguments. Local commands run without a shell. Only HTTPS and loopback HTTP endpoints are supported; legacy SSE is not supported.

Remote MCP connections offer Manual headers or OAuth authentication. Plugin servers can declare oauth: {} for discovery-based registration, or oauth: { "clientId": "…", "callbackPort": 49321 } for a registered public client. New connections inherit these defaults and use the packaged MCP URL; saved authentication choices take precedence. Unsupported OAuth fields are reported as invalid server configuration. New Plugin connections receive a unique default name. For an enabled, approved OAuth server, Connect and sign in saves the connection and opens the system browser. A cancelled or failed login keeps the connection for Sign in to retry. Packaged client defaults remain editable under Advanced sign-in settings. Other OAuth connections use Save, then Sign in. Live Smith uses browser PKCE and the server's discovery metadata. Servers with dynamic client registration choose a local callback port automatically; a later sign-in reuses that registration and port. If a server requires an existing public client, enter its client ID and registered callback port. Its redirect URI must be http://127.0.0.1:PORT/mcp/oauth/callback. Confidential clients and hosted client-ID metadata registration are not configured here. An occupied callback port reports an error; close its other listener or sign out to discard a dynamic registration before trying again.

Saved OAuth accounts can refresh during MCP use; ordinary discovery never opens a browser. Sign out cancels pending sign-in, removes local credentials, and closes active clients. A manual Authorization header cannot be combined with OAuth. Other configured resource headers remain available for workspace routing. Changing the server, package, routing parameters, authentication configuration, or enabled state requires a new sign-in. Unrelated creative Plugin parameters preserve the account. OAuth tokens and registration secrets stay in private local storage, scoped to the exact Connection, and never enter model or dialog state.

Credentials remain write-only environment variables or HTTP headers, never model tool arguments. Plugin connections bind to the exact installed package and server; direct connections bind to their exact launch configuration or URL. Changing that target requires entering credentials again. Direct local MCP connections have separate audio-input and MIDI-output grants. Disabling or removing a connection closes its active MCP clients and preserves saved Session artifacts.

Attachments can be dropped anywhere in the chat surface or pasted into the composer, including while a response is running. Follow-ups → Steer submits the message and its files to the current task at its next safe boundary; Queue keeps them together for a new request after the current response. Later files stay with the next composer draft. Images appear as thumbnails that open a larger preview when clicked; audio includes playback and seek controls. Select another file's name to open it with the system's default application. External applications receive a temporary copy, so their edits do not replace the saved chat attachment. Text and code copies open as .txt files. Select excerpt opens an audio waveform with start/end times and loop audition. Choose Use whole file or Use selected excerpt; the Next request list shows the exact files and excerpt ranges being sent. WAV selection copies whole samples without changing encoding, sample rate or channels. MP3 selection requires Export selection as WAV and attach, which explicitly produces 16-bit PCM WAV at the original sample rate and channel count. MP3 preview/export depends on the browser decoder, has a 30-second decoding deadline and a 128 MiB decoded-sample budget. Files above that decoding budget can still be played and sent intact; export a shorter WAV in an audio editor to create an excerpt.

The original remains saved in the Session. Use original file restores it to the next request; removing an unsent selection undoes that selection and restores the previous draft file. Use again on a history attachment creates a fresh reference to a saved copy without uploading the source again. Removing a history copy leaves its saved source unchanged. Selection and reuse share pending quotas and leave files already captured by Send, Steer or Queue unchanged.

Supported categories include:

Category	Formats and handling
Text and code	UTF-8 and UTF-16 text with any filename extension, including Markdown, CSV/TSV, JSON, YAML, XML, HTML, configuration files, logs, subtitles, and source code. Content is read as inert text.
Documents and tables	PDF; DOCX, XLSX, PPTX; RTF; ODT, ODS, and ODP. Non-PDF documents are extracted as text.
MIDI	MID and MIDI Standard MIDI Files with format 0, 1, or 2 and PPQN timing. Context retains tracks, channels, note timing and velocity, tempo, meter, instruments, and control events.
Images	PNG, JPEG, and WebP are retained directly. Additional browser-decodable images, including GIF, BMP, SVG, AVIF, TIFF, and HEIC/HEIF, are converted to a static PNG.
Audio	WAV and MP3 retain their original bytes, sample rate, bit depth, and channels. Other audio formats require an explicit WAV/MP3 export before attaching; Live Smith does not automatically transcode audio.

Browser conversion depends on the codecs available in the Ableton window. A conversion failure names the file and leaves other files in the batch available to add. Converted images use the new PNG filename and bytes; the source file is not changed. Animated images contribute one static frame. MIDI uses at most 8 MiB; its context is symbolic music data and does not require an audio model. SMPTE-timed MIDI, RMID, and MIDI 2 UMP are not supported. Image, native PDF, and audio use depends on the selected model and connection; attaching a file does not guarantee it can be sent to every model. Model loading consumes provider-returned input modalities and MIME support for both Direct API and subscription Profiles. Raw provider evidence remains visible, while a usable input is marked Supported only when the provider covers Live Smith's concrete formats and the selected protocol can encode them; missing or coarse-only evidence stays unverified. Provider-reported video capability is shown, but Live Smith does not yet accept video attachments. Use paste or drag-and-drop rather than a system file picker. Each Session may hold up to four pending files and 256 MiB total. Images are limited to 5 MiB each; their subtotal is 16 MiB. Documents are limited to 20 MiB total; audio to two files, 128 MiB each, and 15 minutes each. Image-conversion source files are limited to 20 MiB. Model requests have a separate 128 MiB inline binary budget, including history. Provider limits can be lower; if a request is too large, select a shorter audio excerpt or fewer files. Original audio is not automatically compressed or truncated. Locally extracted context is bounded and labels truncation. Modern Office and OpenDocument extraction uses bundled officeparser and SheetJS libraries. No additional runtime or local service is required. Parsing runs in workers with a 30-second deadline and a V8 heap limit. RTF retains a bounded built-in text reader. Spreadsheets retain sheet names, coordinates, stored values and formulas; missing formula caches are marked unavailable. Explicitly hidden Word runs, PPTX slides/shapes, ODP pages, and XLSX sheets/rows/columns are omitted, along with explicitly hidden ODS sheets and collapsed or filtered rows and columns. ODS sheet and ODP page visibility includes referenced and inherited styles. Office formatting, embedded images, and chart geometry are not rendered by local text extraction. Legacy DOC/XLS/PPT, XLSB, and video frames are not accepted as model context; historical legacy Office references remain in saved Sessions and are marked unsupported during context assembly; WebM/MP4 audio also requires an explicit WAV/MP3 export. The Extensions SDK does not expose selected Audio Clip, Sample, or Simpler source bytes to extensions, so export or locate the source file and drop it into the chat when you need the original file as an attachment. During that send, a compatible audio attachment can also be used as the source for an Arrangement, Session, or Take Lane Audio Clip, a Simpler sample, or a Drum Rack pad. When a confirmed plan first uses it, Live Smith copies the file into the Live Project and uses Live's managed copy. The source locator expires when the send ends. For compatible audio models, the agent can instead call read_arrangement_audio to read an isolated Arrangement Audio Clip range without creating a saved attachment. This sends a temporary pre-effects render for the current request; Session View Clips and the track device chain are not included. That temporary render is model input only and is not an attachment SampleSource. Saved Session audio results follow the same verified model-capability gate: the agent can listen to one exact local asset when asked, while remote-only Suno players and their URLs are never sent to the model.

Deleting a Session removes its private chat attachments, but it does not remove audio already imported into the Live Project. If import succeeds and post-import validation or a later Live action stops the plan, an unused project copy may remain because the beta SDK does not expose deletion or rollback for imported files.

Queue and Steer are configured under Settings → App → Conversation & Display. The same section can show or hide the compact context-window indicator in the composer. Queued follow-ups belong to the open window; Live Smith warns before closing with pending work.

Privacy

Profiles, Integration Connections, Sessions, attachments, imported Skills, and installed Plugin packages and private Plugin data are stored locally. Prompts, the active Session's saved creative brief, relevant Live context, selected Skill guidance, supported attachment content, any Arrangement audio range read by the agent, and any saved Session audio the user asks an audio-capable model to hear are sent to the model provider you choose.

Direct API keys are stored in local Profile settings as plain text. A separate private local credential file stores OAuth credentials under exact Profile and provider identities. Saving a connection keeps only the provider selected by that Profile; Direct API and Profile deletion clear that Profile's OAuth credentials. Do not commit, share, or cloud-sync either storage location.

An enabled MCP server receives the arguments declared by its tool. A local server can access anything available to the current operating-system user, subject to that program's own behavior; a remote server receives network traffic at its declared endpoint. Review third-party code, launch commands, and permissions before enabling a server. MCP tool results are treated as untrusted data and cannot grant Live permissions or bypass confirmation.

The selected proxy mode and credential-free Manual proxy URL are stored in the same private local settings file. Proxy usernames and passwords are not accepted in that URL.

Documentation

• Development guide: prerequisites, local setup, validation, packaging, and development data.
• Model connections: connection requirements, model settings, capability evidence, and protocol limits.
• Architecture: module responsibilities, data flow, and safety boundaries.
• Contributor guide: working conventions and required checks.
• Third-party notices.