Eclipse Dispatch MCP

Eclipse Dispatch MCP

Powered by Cyan Eclipse Technologies — https://cetech.ai

Eclipse Dispatch — the newsletter layer above the platforms: it drafts once and syndicates the drafts to every sanctioned platform, unifies the analytics, and watches your own subscriptions like a portfolio. It never publishes without approval, and it never scrapes.

Why this exists

A publisher multi-homed across beehiiv, Ghost, and Substack writes once and re-enters everywhere, then reads three dashboards with three different definitions of an "open." Each platform's own tooling stops at its own wall: Substack's official MCP is read-only analytics gated to Bestseller admins; beehiiv's official MCP drafts into beehiiv only. Dispatch is the cross-platform layer none of them will build — one canonical draft, per-platform drafts via sanctioned routes, analytics unified without silent blending, and review-first as a property of the code, not a policy memo.

Status — honest labels

v0.1.0, phases D0 + D1-core. Bench pending — no certification claimed. Local stdio transport, Python stdlib only. Adapters run against shipped fixtures and local stub HTTP servers shaped like the real APIs — nothing in this build contacts a platform. Live wiring is owner-gated Wave 2: the beehiiv signup is an open owner item, and no Ghost instance runs here. The reader face (spec §5, phase D2 — subscription portfolio, renewal radar, engagement truth) is not in v0.1; neither is the §4.6 unofficial-endpoint module, which the spec recommends not building in v1 at all.

Tools (6)

Tool	What it does
draft_create	Create the canonical draft (title, body, subtitle, canonical_url pointing home, tags, audience, per-platform variants). Versioned and hash-bound: returns the content_hash approval will bind to. Re-creating an existing draft_id makes a new version and resets approval to none.
draft_preview	The review artifact: per-platform rendered previews plus every adaptation diff vs the canonical draft — variant overrides and platform constraints (Substack = paste pack; beehiiv = draft-only, publish in-app).
draft_approve	Record owner approval; returns a token bound to the content hash.
syndicate_drafts	Fan the approved draft out: beehiiv/Ghost drafts via the API-shaped adapters; Substack gets a paste pack (local files the owner posts in-app). Requires the approval token and an un-invalidated hash. Receipt per platform.
analytics_report	D0 unified analytics, read-only. Figures normalized by metric key, each carrying platform provenance + the platform's own definition. A combined figure appears only where definitions are identical across sources; otherwise combined: null flagged DEFINITIONS_DIFFER — flagged, never averaged.
receipt_list	Hash-chained receipts (drafts, approvals, syndications, analytics runs) + chain verification.

The invariants (encoded + tested)

• Hash-bound approval. Approval binds to the canonical draft's content hash. Any post-approval edit — a new version, or a silent in-place tamper — invalidates it; syndication recomputes the hash and refuses by name. There is no path from "edited" to "sent" that passes through a stale approval.
• The beehiiv adapter has no publish method. Deliberate, mirroring the platform's own design (its official MCP also stops at drafts; publishing is the publisher's in-app act). A unit test asserts the method's absence. The server exposes no publish tool at all — no auto-publish, at any tier, for any platform.
• The Substack adapter is read-only. No sanctioned write surface exists, so write attempts raise a named refusal (ReadOnlyRefusal: "Substack adapter is READ-ONLY…"). Substack receives paste packs, never writes. The spec's unofficial-endpoint module is not built (spec decision 3 recommendation: not in v1) — its absence changes nothing.
• Analytics never blend silently. Three platforms define "open rate" over three different populations (the fixtures say so, in each definition). Dispatch labels every figure and refuses to average across definitions.
• No scraping. Public RSS of the owner's own publication is consumption as offered; bulk extraction does not exist here (spec §7).

Receipts

Draft creations, preview packs, approvals, per-platform syndications, and analytics runs append hash-chained receipts — append-only and machine-verifiable (verify_receipt_chain); tampering with any entry breaks the chain and is detected by test.

Run it

python3 server.py                      # MCP over stdio, stdlib only
python3 smoke.py                       # end-to-end smoke over real stdio (temp state, fixture adapters)
python3 -m unittest discover -s tests  # 15 tests

State persists in data/state.json beside the server (override with ECLIPSE_DISPATCH_STATE); paste packs land in data/paste-packs/. data/ is gitignored. Adapter base URLs can be pointed at stub (or, in Wave 2, real) endpoints via ECLIPSE_DISPATCH_BEEHIIV_URL / ECLIPSE_DISPATCH_GHOST_URL; with none set, in-process fixture transports answer.

Fixtures

fixtures/ ships per-platform analytics (each metric with the platform's own name and definition — the beehiiv and Ghost revenue figures deliberately use different definitions, period gross vs MRR, so the no-blending rule has something honest to catch) and substack_feed.xml, the owner's own publication RSS in fixture form. Stub servers shaped like the beehiiv v2 API and the Ghost Admin API live in adapters.py and are exercised over real local HTTP by the tests.

Roadmap (per spec)

• D2 — reader face: subscription portfolio graph, renewal radar, engagement truth, cancel prep (own data only, strictest data tier).
• D3 — Substack depth as the sanctioned surface grows — design waits for sanctioned writes; no scraping milestone exists.
• Wave 2 (owner-gated) — live beehiiv wiring after the owner's signup; live Ghost against a real instance.

License / attribution

A Cyan Eclipse Technologies product. The "Powered by Cyan Eclipse Technologies" attribution travels with it.