Almost Impossible API
Our published work, team and pages, readable by software. A REST API with an OpenAPI spec, an MCP server and an NLWeb endpoint, all read-only. There is no API key and nothing to sign up for.
Base URL https://www.almostimpossible.agency/api/v1 · Spec /openapi.json · MCP /mcp
Endpoints
- GET
/api/v1/workPublished case studies. ?q filters by title, client or summary; ?limit (1 to 20) sets the page size and ?cursor takes next_cursor from the previous page. - GET
/api/v1/work/{slug}One case study: client, year, services, summary and a short TL;DR. - GET
/api/v1/teamThe team: names, roles and short bios. No email addresses. - GET
/ask?query=…A question in plain language, answered with schema.org items from the site (NLWeb).
curl "https://www.almostimpossible.agency/api/v1/work?q=brand&limit=5"Responses are JSON and cached for up to five minutes. Errors use RFC 9457 problem details (application/problem+json) with type, title, status and detail. Every endpoint is described in the OpenAPI 3.1 spec, which you can load straight into a function-calling model or an API client.
Sandbox: there is no separate test environment because none is needed. Every endpoint is read-only, so the production API is safe to call while you build, with nothing to break and no data to clean up.
Versioning: the version is in the path, /api/v1. Within v1, changes only add fields or endpoints. A breaking change would ship as /api/v2, announced here first; v1 would then send Deprecation and Sunset headers for at least 90 days before it is removed.
MCP server
https://www.almostimpossible.agency/mcp speaks the Model Context Protocol over Streamable HTTP, revisions 2025-03-26 to 2026-07-28. It needs no sign-in. Its registry name is agency.almostimpossible/site and its server card is at /mcp/server-card. Add it to an MCP client:
{
"mcpServers": {
"almost-impossible": {
"type": "http",
"url": "https://www.almostimpossible.agency/mcp"
}
}
}Every tool is read-only:
search_siteSearch the website.get_pageRead a page as Markdown.get_agency_overviewAgency overview.list_workList case studies.get_workGet one case study.list_teamList the team.
Each page of the site is also a resource, as Markdown, alongside llms.txt.
Ask the site a question
/ask follows NLWeb v0.55. Send a question and get schema.org items back: pages, case studies, news and articles. Ask for text/event-stream to stream the results.
curl -X POST https://www.almostimpossible.agency/ask \
-H 'Content-Type: application/json' \
-d '{"query": {"text": "rebrand work for banks"}}'WebMCP tools
In a browser that supports WebMCP, every page registers tools for the agent already on it: list and read case studies, list the team, and create_enquiry_draft, which fills in the enquiry form for the person to review and send. Nothing is sent without them.
Machine-readable files
- /llms.txt: The agency, its work and how an agent should engage, in one file.
- /openapi.json: The REST API, as OpenAPI 3.1.
- /schemamap.xml: NLWeb schema map, pointing at the JSONL feed of case studies.
- /.well-known/ai-catalog.json: Every agent surface in one catalogue.
- /sitemap.xml: Every public page.
Questions, or a use we should support? Tell us through the enquiry form.