Integration

CompactifAI integration

CompactifAI by Multiverse Computing is an OpenAI-compatible inference API for compressed open-weight language models (its own Carina and Quasar models, GLM and Qwen) and a compressed Whisper speech-to-text model. List the models and their capabilities, send chat completions, text completions, Responses API calls (stored responses can be read back) or Anthropic-format messages, count message tokens for free, transcribe audio, check a text for the model's watermark, run a web search, and run 24-hour batch jobs from an uploaded JSONL file. Usage is billed pay-as-you-go per input, cached input and output token (transcription per audio second); prompts and completions are not retained, except files you upload and responses created with store true, which are kept on US servers. Streaming is not exposed, and neither are embeddings: no listed model serves them. Authentication is the account's CompactifAI API key from the dashboard, sent as a Bearer token in the Authorization header.

Connect CompactifAI How integrations work

Ask it in chat

Say what should happen.
The workflow builds itself.

When this happens, do that in CompactifAI.Built from the 17 CompactifAI tools below, on the connection you add once.

What you can do with CompactifAI

Each of these is one operation on one endpoint, available both as a workflow step and as a tool an AI agent can call on its own. Nothing bundles several operations into one action, so what a run did stays readable afterwards.

Each badge shows the default agent permission of an endpoint (allowed: an agent calls it without asking; ask first: you are asked before an agent calls it, when the agent or chat asks permission for sensitive actions; never: no agent may call it), which every user can change endpoint by endpoint in their credential settings.

Most endpoints run on the free plan. A few, mainly the ones that publish on your behalf, need a paid plan.

Batch, 4 operations

cancel_batchCancel batchPOSTAgents: ask firstCancels an in-progress batch job: stops it from running the remaining requests. Returns the batch, usually with status cancelling.
create_batchCreate batchPOSTAgents: ask firstStarts a batch job that runs every request line of an uploaded JSONL file (upload_file with purpose batch) against one endpoint within 24 hours; poll get_batch until completed, then download output_file_id.
get_batchGet batchGETAgents: ask firstReturns a batch job's live state by id: status (validating, in_progress, finalizing, completed, failed, expired, cancelling or cancelled), request counts and the output and error file ids.
list_batchesList batchesGETAgents: ask firstLists the account's batch jobs with live status and request counts. Paged: pass last_id as after for the next page; limit 1 to 100 (vendor default 20).

Model, 2 operations

get_modelGet modelGETAgents: ask firstReturns one model's details and capabilities by its id (as list_models returns it): owner, creation time and which input types and tools it supports.
list_modelsList modelsGETAgents: ask firstLists the available models with their capabilities (image, video, audio, transcription, tool calling, and which of chat completions, responses, messages and embeddings each serves); pass an id as model to the other tools.

Response, 2 operations

create_responseCreate responsePOSTAgents: ask firstGenerates a model response in the OpenAI Responses format; with store true (the vendor default) it is kept on US servers so get_response can read it and previous_response_id can continue it. Billed per token.
get_responseGet responseGETAgents: ask firstReturns a stored response by its id (from create_response called with store true), with its output items, status and token usage. Only the account that created it can read it.

Chat completion, 1 operation

create_chat_completionCreate chat completionPOSTAgents: ask firstGenerates a chat reply from a model (OpenAI Chat Completions format), with optional tool calling, structured output, image, audio or video input and reasoning controls. Billed per input and output token.

Completion, 1 operation

create_completionCreate completionPOSTAgents: ask firstGenerates a text continuation of a prompt (legacy OpenAI Completions format). Billed per input and output token; max_tokens caps the output (vendor default 16).

File, 1 operation

upload_fileUpload filePOSTAgents: ask firstUploads a file and returns its id: a JSONL file with purpose batch feeds create_batch, a user_data file can be passed to create_response as file_id. Uploaded files are stored on US servers.

File content, 1 operation

download_file_contentDownload file contentGETAgents: ask firstDownloads a file's raw contents and stores them as a file, by id: an uploaded file, or a batch's output_file_id or error_file_id from get_batch.

Message, 1 operation

create_messageCreate messagePOSTAgents: ask firstGenerates a reply in the Anthropic Messages format (model, messages and max_tokens are required), for models whose capabilities include messages. Billed per input and output token.

Message token count, 1 operation

count_message_tokensCount message tokensPOSTAgents: ask firstCounts the input tokens an Anthropic-format message would use, without generating anything. Free (no tokens are billed), but it counts against the request rate limit.

Transcription, 1 operation

transcribe_audioTranscribe audioPOSTAgents: ask firstTranscribes an audio file (flac, mp3, mp4, mpeg, mpga, m4a, ogg, wav or webm) to text with a transcription model such as cai-whisper-large-v3-turbo-slim. Billed per audio second, one minute minimum.

Watermark, 1 operation

detect_watermarkDetect watermarkPOSTAgents: ask firstChecks whether a text was generated by a model with text watermarking on; returns watermarked and a confidence (none, low, medium or high). Not detected does not prove a person wrote it.

Web search, 1 operation

search_webSearch webPOSTAgents: ask firstSearches the web and returns up to count results (1 to 50, vendor default 10), each with title, content snippet, link, media, icon and publish date.

Use CompactifAI from Claude Code, Cursor or Codex

Connect CompactifAI in LiveContext, create an API key in Settings, then add the LiveContext MCP server to your client. Its CompactifAI operations become tools the client can call, next to every other integration you connected.

Claude Code

claude mcp add --transport http livecontext https://livecontext.ai/mcp --header "X-API-Key: YOUR_API_KEY"

Codex (~/.codex/config.toml)

[mcp_servers.livecontext]
url = "https://livecontext.ai/mcp"

[mcp_servers.livecontext.http_headers]
"X-API-Key" = "YOUR_API_KEY"

Cursor, Claude Desktop and other JSON configs

{
  "mcpServers": {
    "livecontext": {
      "type": "http",
      "url": "https://livecontext.ai/mcp",
      "headers": {
        "X-API-Key": "YOUR_API_KEY"
      }
    }
  }
}

Set up the MCP server step by step

How to connect CompactifAI

Create your workspace
Sign up free. Nothing to install.
Connect CompactifAI once
Add your CompactifAI connection in Settings. Every tool from this integration reuses it, and secrets are encrypted at rest and visible only to you.
Describe the job in chat
The workflow builds itself in front of you and picks the endpoints it needs. Drag the steps yourself on the canvas whenever you would rather be explicit.