Consuming OpenStream

Public tier needs no key. A key (os_live_… / os_test_…) via Authorization: Bearer lifts you to the customer tier. Every response carries a dataset version and a per-response citation object.

GET /api/v1/answers/cheapest-nbn-100            # small, answer-shaped JSON (agents: start here)
GET /api/v1/answers                                # index of answer slugs
GET /api/v1/plans?vertical=nbn&speed=100&connection=fixed_line&sort=price&limit=20   # paginated: limit/offset, total, next_offset
GET /api/v1/plans?vertical=nbn&min_speed=100&verbosity=full
GET /api/v1/plans/{id}?as_of=2026-07-10        # point-in-time (valid-time)
GET /api/v1/plans/{id}/history                 # price-over-time
GET /api/v1/changes?since=2026-07-01T00:00:00Z # diff-first change feed
GET /api/v1/margin?product_id={id}             # open margin, or an explicit reason
GET /api/v1/zones?postcode=3000                # postcode -> energy zone
GET /api/v1/calculator?product_id={id}&annual_kwh=4000
GET /api/v1/openapi                            # OpenAPI 3.1 (wire into a Custom GPT)

Remote MCP over Streamable HTTP at https://openstream.com.au/api/mcp. No key needed for the public tier (rate limited per IP); add Authorization: Bearer os_live_… for your key's limits. Works in Claude Code, Claude Desktop and claude.ai (custom connector), ChatGPT (developer mode and deep research via search/fetch), Codex, Cursor and any Streamable-HTTP client.

# Claude Code
claude mcp add --transport http openstream https://openstream.com.au/api/mcp
# with a key
claude mcp add --transport http openstream https://openstream.com.au/api/mcp --header "Authorization: Bearer os_live_..."

# OpenAI Codex CLI
codex mcp add openstream --url https://openstream.com.au/api/mcp

# .mcp.json / Cursor / Windsurf
{ "mcpServers": { "openstream": { "type": "http", "url": "https://openstream.com.au/api/mcp" } } }

# Claude Desktop / claude.ai: Settings → Connectors → Add custom connector → paste the URL
# ChatGPT: Settings → Connectors (developer mode) → Add → MCP server URL

Tools: check_address (NBN Co lookup for a street address: connection type, eligible plans, a resolved link) · cheapest_electricity (every published plan on a postcode's network, costed at a stated usage) · recommend_electricity (the site's electricity picks for a postcode and bill, each with a plan page) · electricity_signup_link (tracked hand-off to the retailer) · electricity_network (reference price, cheapest, median, overnight moves) · find_plans · cheapest_plan · compare_plans · get_plan · plan_history · plan_changes · lookup_provider · search · fetch. Every result carries price_last_confirmed_at and a permanent dated page_url to cite. Try in chat: "cheapest NBN 100 today?", "compare Aussie Broadband and Superloop 1000", "what changed at TPG this week?".

• Units inside the payload (c/kWh), gst_inclusive on every money field, ISO dates.
• Structured refusals that say what's missing: {"error":"zone_required","how_to_resolve":"call /zones?postcode="}.
• Verbosity budget: summary vs full.
• Every response reproducible against its dataset_version.

Server-rendered, permanently-dated, citable pages: one per plan per day of state, plus a delta log per day. schema.org Offer markup, llms.txt, per-fact anchors, and the provenance triple printed on every page. This is the layer ChatGPT search and other indexers cite directly - no key required.