cypher-mcp

cypher-mcp

Monetized graph answers over Bitcoin Lightning. A Tollbooth-DPYC Operator that sells priced answers from a graph — operator-authored, parameterized, named Cypher query templates — not raw database access.

The OSS Neo4j MCP servers are a free pipe to a database the operator separately expenses (Neo4j invoice) and separately bills for (Stripe stack). cypher-mcp collapses infrastructure cost, application billing, and access control into one metered MCP surface: the operator stops selling seats and starts selling answers, settled in sats, with no Stripe stack and no Neo4j invoice passed to the customer.

Live at https://cypher-mcp.fastmcp.app/mcp.

How it works

Two planes, cleanly separated:

• Settlement / identity plane — Neon + Nostr + Tollbooth, shared with every other DPYC operator. It never knows a graph exists.
• Value-delivery plane — Bolt → Neo4j, sealed inside the tool handlers (graph.py), opened and torn down per request (no pool). The backing store is an implementation detail invisible to patrons, so it can be swapped without changing the MCP surface.

Three conceptual layers:

Layer	Where	What
L1 — primitive	code (shipped once)	execute_query_by_key(key, params) + catalog CRUD
L2 — catalog	Neon query_catalog	key → vetted Cypher template + param schema (what's possible)
L3 — pricing	Tollbooth pricing model	what's sold, named, priced

Status: live, with named tools. L1 + L2 are shipped and serving real, billed answers — an operator adopted by an Authority, AuraDB-backed, with refund-on-raise verified end to end. Named tools are now live (L3 tool synthesis): an operator can publish_tool(key) to project a query as a first-class, typed tool (e.g. cypher_which_service_handles(keyword)), priced individually in Pricing Studio. Per-key parametric pricing formulas remain a later step — see CHANGELOG.md.

Tools

Patron (priced):

• cypher_execute_query_by_key(key, params) — run a published query. You supply only parameters; the operator owns the query text, and parameters bind as Cypher $params (never string-interpolated). You're charged only for a delivered answer — an unknown key, invalid parameters, or a failed query rolls back the debit (refund-on-raise).

Free: cypher_public_factory_stats — aggregate counts only (no titles, paths or npubs), cached in-process, for the unauthenticated landing pages of the read-only Cypher Lab Notebook front end in frontend/.

Plus the standard DPYC surface from the wheel: cypher_check_balance, cypher_check_price, cypher_purchase_credits, Secure Courier, npub proof, Oracle delegation, account statements, constraint-based dynamic pricing, etc.

Operator-only (restricted, unpriced) — the authoring plane:

• cypher_create_query / cypher_update_query / cypher_get_query / cypher_list_queries / cypher_delete_query
• cypher_publish_tool(key) / cypher_unpublish_tool(key) — project a query as a first-class, typed named tool (or retire it). See Named tools below.

Raw Cypher never reaches patrons — only named, parameterized templates.

Authoring in Neo4j's own UI — no second editor to learn

A DB-analyst refines templates in Neo4j Browser, not in a bespoke tool. cypher_get_query(key) returns the template plus an edit_url — a deep link into the hosted Neo4j Browser (browser.neo4j.io) that pre-targets the operator's AuraDB and loads the statement in edit mode (cmd=edit). Refine it with autocomplete, run-to-test, visualization, and PROFILE; then save it back with cypher_update_query. The template is just a string flowing over MCP — no files, no desktop, no exported bundles. The link carries the DBMS URI but never the password, and get_query is operator-only, so neither ever reaches a patron.

Named tools — publish a query as its own tool

cypher_publish_tool(key) projects a catalog query as a first-class MCP tool named cypher_<key> whose flat, typed parameters come from the query's schema — e.g. cypher_which_service_handles(keyword). An agent then calls it by name with typed params instead of execute_query_by_key(key, params). Internally every named tool funnels through the same shared executor (look the stored Cypher up by key → parameter-bind → run), so there is one code path and the same refund-on-raise guarantee. A published tool registers unpriced — it appears in Pricing Studio like any new tool; set its price there (calls return "not priced yet (TBD)" until you do). cypher_unpublish_tool(key) retires the tool (the query itself stays, still runnable by key). Published tools survive restarts (re-materialized from the catalog); reconnect to observe tool-list changes.

The synthesis machinery is a generic tollbooth-dpyc primitive (register_dynamic_tool) — named Cypher queries are one realization; the same primitive can back a synthesized tool with a REST call, SQL, or a stored prompt.

Onboarding (operator)

1. Request adoption from an Authority (cypher_request_adoption, signed with the operator's key). Adoption is the only way in: once the Authority's owner approves the request, the Authority provisions an isolated Neon tenant automatically. cypher_adoption_status reports progress.
2. Deliver operator secrets via Secure Courier (cypher_request_credential_channel, service cypher-operator):
   • neo4j_uri, neo4j_user, neo4j_password — the graph store (e.g. a Neo4j AuraDB instance)
   • btcpay_host, btcpay_api_key, btcpay_store_id — Lightning
3. Author your catalog with cypher_create_query; patrons then execute by key.

Develop

pip install -e ".[dev]"
pytest -v
python -m cypher_mcp.server   # needs TOLLBOOTH_NOSTR_OPERATOR_NSEC

Deployed on Prefect Horizon (FastMCP runtime); see fastmcp.json. The only env var required to boot is TOLLBOOTH_NOSTR_OPERATOR_NSEC — every other secret arrives via Secure Courier.

---

DPYC Ecosystem

Repo	Role
tollbooth-dpyc	Python SDK — vault, auth, pricing, payments, Nostr identity
dpyc-community	Governance registry, membership, advisories, threat model
dpyc-oracle	Community concierge — free onboarding help and membership lookup
tollbooth-authority	Certification backbone — Schnorr-signed purchase order certificates
tollbooth-sample	Sample Operator — canonical template for new MCP services
tollbooth-pricing-studio	iOS app — visual pricing-model editor and operator console
schwab-mcp	Charles Schwab brokerage data (operational example)
thebrain-mcp	TheBrain personal knowledge graph (operational example)
excalibur-mcp	X/Twitter posting (operational example)
taxsort-mcp	Tax classification + Cloudflare Pages UI (operational example)
goodearth-mcp	Region-scoped farm climate analytics (operational example)
beesknees-mcp	Multiplayer race to the queen (operational example)
roastify-mcp	Roastify coffee catalog and packaging artwork (operational example)
ChartRemotely-mcp	Remote desktop charting via a patron's own agent (operational example)
tollbooth-web	Shared browser SDK for operator front ends
tollbooth-oauth2-collector	OAuth2 callback handler — shared advocate service
stablecoin.myshopify.com	DPYC merch and Austrian economics

Prior Art & Attribution

The methods, algorithms, and implementations contained in this repository may represent original work by Lonnie VanZandt, first published on June 15, 2026. This public disclosure establishes prior art under U.S. patent law (35 U.S.C. 102).

All use, reproduction, or derivative work must comply with the Apache License 2.0 included in this repository and must provide proper attribution to the original author per the NOTICE file.

How to Attribute

If you use or build upon this work, please include the following in your documentation or source:

Based on original work by Lonnie VanZandt and Claude.ai
Originally published: June 15, 2026
Source: https://github.com/lonniev/cypher-mcp
Licensed under Apache License 2.0

Visit the technologist's virtual cafe for Bitcoin advocates and coffee aficionados at stablecoin.myshopify.com.

Patent Notice

The author reserves all rights to seek patent protection for the novel methods and systems described herein. Public disclosure of this work establishes a priority date of June 15, 2026. Under the America Invents Act, the author retains a one-year grace period from the date of first public disclosure to file patent applications.

Note to potential filers: This public repository and its full Git history serve as evidence of prior art. Any patent application covering substantially similar methods filed after the publication date of this repository may be subject to invalidation under 35 U.S.C. 102(a).

Further Reading

The Phantom Tollbooth on the Lightning Turnpike — the full story of how we're monetizing the monetization of AI APIs, and then fading to the background.

Software Factory intent vocabulary (DPYC agents)

cypher-mcp is the graph substrate for the DPYC Software Factory: its agents (Porter = Service Desk, Journeyman = Engineering) read and write an institutional-memory graph here without ever holding a raw Cypher tool. They call a fixed vocabulary of operator-authored, parameterized templates — write templates and read templates — published as typed cypher_<key> tools.

• Write templates capture the memory: record_triage, note_rejection, route_rejection (anti-ping-pong: bounce a rejection back to its origin), record_scope, link_root_cause, link_pr, assert_rationale, bind_rationale_to_symbol, anchor_symbol (post-edit grep-scoping anchor), upsert_capability, assert_invariant, upsert_patent_element, register_service, and more.
• Read templates are the intention-service forward map — the RAG surface that resolves "which service handles X / why does this capability exist" in one query instead of a twelve-repo grep: which_service_handles, explain_capability, what_realizes_capability, capability_patents, context_pack (the flagship grep-scoping bundle), list_capabilities, list_invariants, list_patent_elements, routing_history, recent_activity, factory_resolution_stats, and their provenance readers.

assert_rationale hard-codes provenance:'llm-inferred-unverified' as a Cypher literal, so an agent key can never claim authoritative provenance.

Start unlimited, limit later. Seeding just authors + publishes the tools; they then appear in Pricing Studio unpriced. Price them and any funded patron can call — the two agents just purchase api_sats like any patron. When you want to restrict calls to the two, add — in a Pricing Studio session — a json_expression allow-list on patron.npub per tool (Constraint Engine, not a bespoke ACL): listed npub allowed, everyone else denied by default, runtime-mutable with no deploy. Containment is also by balance — fund the Porter thin and the Journeyman thicker; a drained agent stops writing the graph but still triages on GitHub.

Seeding (operator, idempotent). scripts/seed_factory_vocabulary.py authors + publishes the vocabulary. --dry-run needs no server or nsec and also prints the exact allow-list to replicate in Pricing Studio when you limit access:

python scripts/seed_factory_vocabulary.py --dry-run \
    --porter-npub npub1ymg... --journeyman-npub npub1m5q...

Apply against the live operator (the operator nsec, read from OPERATOR_NSEC or a hidden prompt, is used transiently to sign each request (NIP-98) — never logged or written to disk). Add --gate (with both npubs) to also apply the per-npub allow-list programmatically instead of in Pricing Studio:

python scripts/seed_factory_vocabulary.py \
    --url https://cypher-mcp.fastmcp.app/mcp --operator-npub npub1fuhq0... \
    --porter-npub npub1ymg... --journeyman-npub npub1m5q...

The vocabulary itself lives in scripts/factory_vocabulary.py (data, no I/O) and is covered by tests/test_factory_vocabulary.py. Pipeline wiring (agents signing and calling these tools), NIP-05 publication, and funding are downstream of this substrate.

Trademarks

DPYC™, Tollbooth DPYC™, and Don't Pester Your Customer™ are trademarks of Lonnie VanZandt. See TRADEMARKS.md in the dpyc-community repository for usage guidelines.

License

Apache License 2.0 — see LICENSE and NOTICE for details.

---

Because in the end, the tollbooth was never the destination. It was always just the beginning of the journey.