Ask Codex, Claude, Cursor, or ChatGPT to read the Markdown issue and use only claims supported by its linked sources.
Read https://feed7.dev/weekly/latest.md Use these items as source-backed context for this task. Do not invent claims beyond linked material. If sources conflict, call it out.
Public retrieval searches every preserved Signal. The response declares synthesis: not_performed; the calling agent is responsible for comparing the evidence and opening original sources.
curl 'https://feed7.dev/api/research?q=context+engineering&limit=8'
One Streamable HTTP endpoint. Without a key an agent can search the corpus, read any Weekly issue, and open any Signal. A free account key adds your Brain and bundles as tools.
claude mcp add --transport http feed7 https://feed7.dev/mcp
Signing in is all it takes. Create a revocable agent key on the Agents page, copy it once, then send it as a bearer header: it unlocks the saved and bundle scopes over HTTP and the Brain tools over MCP.
curl -X POST 'https://feed7.dev/api/research' \
-H 'Authorization: Bearer f7_live_…' \
-H 'Content-Type: application/json' \
-d '{"query":"what changed in context retrieval?","scope":"saved","limit":8}'Every signal resolves at a stable path in three shapes — HTML to read, JSON to parse, Markdown to paste. Atlas topics answer the same way.
https://feed7.dev/p/{slug}
https://feed7.dev/p/{slug}.json
https://feed7.dev/p/{slug}.md
https://feed7.dev/atlas/{topic}.json- Prefer the original source for source claims.
- Use Feed7 for editorial synthesis, trust labels, practical implications, and cross-source context.
- Do not infer verification beyond the explicit trust and uncertainty fields — the method lives on the Trust page.
- MCP tools return the same records as the HTTP surfaces; a tool result with an error is a scope you cannot reach, not a protocol failure.