Search · scrape · summarise
Search results and pages, turned into a summary you can cite.
A FastAPI service that scrapes search results and web pages, then synthesises an executive summary with any of 54 model providers, from a local Ollama to the hosted ones. API only, no UI.
POST /v1/search
{
"queries": [
{"query": "widget latency causes", "max_results": 10},
{"query": "dispatch queue sharding", "site": "reddit.com"}
],
"fetch_pages": true,
"max_pages": 3,
"additional_notes": "The reader is an SRE deciding whether to shard."
}
# one bad query never fails the batch: every item has its own statusHow it works
Three operations, one summary step, and no key of its own.
Every third-party credential is stored in keyring and resolved per request. Local model runtimes need no credential and work without keyring at all.
Search
A batch of queries, each with its own notes and optional site filter. By default the summary is built from titles and snippets; fetch_pages also reads the top pages.
Scrape
A batch of URLs to clean extracted text. render_js is auto, escalating from plain HTTP to a headless browser when the page looks client-rendered, or always, or never.
Summarise
Text in, summary out, with the model you name as provider:model. A direct hook for an MCP server.
The API
Seven routes, and interactive docs at /docs.
The reference is docs/api.md; the provider catalogue is docs/providers.md.
Search
Batch queries to results, an optional deep fetch, and a summary.
Scrape
Batch URLs to clean extracted text and a summary.
Summarise
Text in, summary out.
Models
What is reachable right now, with per-model capabilities.
Jobs
Poll, list and cancel work started in background mode.
Readiness
200 when a browser can launch and a provider is configured; otherwise 503 naming which.
- Install.
make install - Point it at your keyring.
cp .env.example .env - Serve it; the interactive docs are at
/docs.make run
Where provider keys live, and how they are resolved per caller, is in docs/keyring.md.
What to know first
What to know before you lean on it.
Where do provider keys live?
Not here. Every third-party credential is stored in keyring and resolved per request for the person the request is made for, so answers are per caller.
What happens when a search engine blocks the request?
That item carries its own error and the batch still answers 200. A page that cannot be fetched simply keeps its snippet: deep fetching is enrichment.
Do background jobs survive a restart?
No. Jobs live in the process that accepted them: they are lost on restart and are not visible to other replicas, so behind a load balancer a client must poll the instance it submitted to.
Is it affiliated with a search engine or a model provider?
No. It is independent software and is not affiliated with Google, or with any model provider it can call.
Docs
Written for the person who has to run it.
The HTTP contract
Every route, field and error.
Providers
Which model providers it can summarise with.
Keys and keyring
How a credential is resolved per caller.
Security
What it fetches, and what it refuses to.
Operations
Every setting, and running it for more than one person.
Assistant tools
The operations as an MCP server would offer them.