Geiger Grey · API, MCP and widget
Put the answers where people ask
Any OpenAI client, grounded answers
Every published pipeline serves an OpenAI-compatible chat completions endpoint. Answers stream back with the citations behind them, and a retrieve endpoint returns the passages on their own.
Geiger GreyCitations arrive before the answer does
The first streamed chunk lists every passage the answer can cite, so a client can resolve each [n] as the text arrives, and the last chunk lists the ones it used. A models endpoint names the published version that answered.
Keys scoped to what they're for
Give each key the Retrieve, Chat or MCP scope, limit it to certain pipelines, and set a rate limit and an expiry. The full key is shown once, then only its prefix.
Just the passages, if you'd rather
The retrieve endpoint returns the top chunks for a query with their source, page or timestamp, score and how each was found, and writes no answer.
Grounded answers inside your AI tools
An MCP server puts a pipeline and its knowledge bases in front of MCP clients such as Claude Desktop and Cursor, and only clients holding its API key can connect.
Geiger GreyFour tools, with search and ask on from the start
New servers expose search and ask; switch on get_entity and list_sources when a client needs them. Copy the config for Claude Desktop or Cursor from the server's page, or take the server offline with one switch.
Answers arrive with their sources
The ask tool runs the server's published pipeline and returns the answer followed by its numbered sources, pages and timestamps included.
Scope each server
Each server answers from one pipeline and the knowledge bases you pick, signs clients in with one of your API keys, and exposes only the tools you switch on.
A chat bubble for your own site
Paste one script tag and visitors can ask questions in a chat bubble that answers from your knowledge bases and names every source, on model keys you bring yourself.
Geiger GreyCited answers where your visitors already are
The widget streams each answer from one pipeline's published version, then lists its sources by title and page or timestamp. It loads only on the domains you allow.
Make it sound like you
Set the title, greeting and placeholder, put the bubble on the left or the right, and hide the Powered by Grey footer.
One script tag, your domains only
Each widget answers from one pipeline, loads only on the hostnames you allow, and can be disabled without touching your site.
Your own model keys
Connect OpenAI, Anthropic, Google, Cohere, Voyage, AssemblyAI, OpenRouter or any OpenAI-compatible server, Ollama and vLLM included. Keys are kept in Vault and only their last four characters are shown.
Usage in units, with a cap you set
Usage counts pages ingested, media minutes, retrievals, tokens and storage. When the monthly soft cap is reached, Grey can alert you only, pause ingestion but keep retrieval, or pause ingestion.