Agiscorecard Web3 Protocol

Web3 Workbench

Ten independent, usable free-beta tools and one hub. Source and generated fixtures are shared; every host serves only its own pages. Existing Agent Delivery Lab remains a separate deployment.

Hub: https://web3.agiscorecard.com/

Development and release

Node.js 22+ and Python 3 with Pillow and DejaVu Sans fonts are required. Browser/offline calculation engines have no npm dependencies; the remote MCP server uses the pinned official SDK.

python3 -m pip install -r requirements-build.txt

npm ci npm test python3 scripts/check_html.py npx --yes wrangler@4 deploy --dry-run

npm test builds 55 HTML pages, 30 worked scenarios, per-host metadata and share images, and the offline bundle, then runs 71 tests. WEB3_BUILD_REVISION pins the actual checked-out commit. Generated dist/, release module and credential-free generated Wrangler config are not committed.

The path-scoped GitHub workflow validates, checks all hostname ownership, binds the existing database, deploys the Worker, and checks every production host and asset against the build. A collision stops before publication. Existing credentials remain in GitHub Secrets. Branch runs validate only; main may deploy. Do not substitute the trigger SHA for the checkout SHA after the stale-main guard.

Only web3_studio_feedback is created/read/updated. No other database table is modified. Optional feedback is constrained enums; IP is used only as the transient rate-limit key. QA samples are explicitly flagged and excluded from demand. Grouped submission counts never represent verified individuals. TTL cleanup occurs on feedback/stats requests and may be delayed during inactivity.

Product limits

Browser inputs are local. No model provider, wallet, live-chain simulation, transfer, subscription or proof verification is connected. Documentation states each tool's own limits. Protocol records are editorial summaries, not complete schemas or compatibility certification. New algorithms mean reproducible combinations of known methods here, not a novel cryptographic or economic theorem.

Research and falsification conditions: portfolio research. Every tool's catalog includes alternatives, scope, commercial hypothesis and a concrete validation condition. Public downloads are limited to the explicit offline bundle and its license; no repository archive is served.

Experience, discovery and sharing (1.1)

Each tool has three runnable scenarios, per-record-group CSV paste, same-tab result comparison, Markdown reports and print/PDF output. Public share links allow only known scenario IDs; private inputs never appear in generated URLs. Clearing/reloading the tab removes in-memory records and comparisons.

Search titles describe actual tasks. Human-readable methods and examples back the structured WebApplication, FAQ and breadcrumb data. No fabricated ratings, usage statistics or rich-result promises. guide.md, tool.json, input.schema.json, llms.txt and llms-full.txt are navigation/integration aids, not special eligibility requirements for AI search. The schema is structural; the engine enforces its additional constraints.

/publish.html on the hub provides ten share cards and optional branded links. Distributed HTML card links use rel="nofollow". No automatic outreach, paid links or third-party endorsements are claimed.

The existing weekly/manual IndexNow workflow validates the key and the canonical page set on each of the 11 hosts before submitting. It never submits on push. node scripts/indexnow.mjs --check validates without submitting. A 200/202 submission response is not evidence of indexing or ranking.

Release smoke verification uses asset-manifest.generated.json to hash-check the exact assets for each host; generated files are excluded from git. See growth implementation record.

MCP (1.2)

Unified endpoint: https://web3.agiscorecard.com/mcp. Each tool host also exposes /mcp, scoped to its own calculator. Streamable HTTP, no key; official SDK 2.0.0 handles 2026-07-28 per-request negotiation and stateless 2025-era compatibility. GET/DELETE return 405; persistent subscriptions are not offered. The existing AGI Scorecard main-domain MCP remains separate.

The hub exposes ten deterministic calculators plus search and fetch. All results preserve the existing algorithms, exact amounts, release revision, methodology citation, date and limits. resources/list exposes public method references, capabilities and fictional inputs. No arbitrary resource fetching or private report publication is implemented. Annotations declare read-only, non-destructive, idempotent, closed-world tools.

Remote MCP arguments are sent to the server, unlike browser/offline inputs. Application code neither persists them nor logs their contents or sends them to a model/source URL. Hosting and the connecting client's own policies still apply. No D1 writes occur in MCP. A separate 120/minute/IP binding limits calls across the 11 hosts; body size is at most 128 KiB including the protocol envelope. Missing rate-limit binding fails closed. Exact Origin validation rejects unknown browser origins; server-side clients without Origin remain supported.

/for-agents.html is the human setup/citation/privacy guide. /.well-known/mcp.json is explicitly a descriptive metadata file, not a standard auto-install mechanism or official registry listing. MCP availability is not evidence of AI search citations.

Tests use the official 2.0 client pinned to 2026-07-28 and the official 1.30 client for 2025-era compatibility. scripts/mcp-smoke.mjs repeats negotiation, tool discovery, resources, all 30 scenarios, citations and host scoping against the deployed endpoint. CI installs the lockfile, retains the stale-main guard and uses existing Cloudflare credentials. No new scheduled workflow is introduced.

Market context and conversion measurement (1.3)

The hub now serves source-stamped /market.html, /briefs.html, three task guides, /api/market, /api/briefs and /updates.xml. Price snapshots refresh on request after five minutes; official metadata after one hour. Stale/unavailable data is explicit and data older than a day is hidden. These are venue references, not executable quotes or verified payments. Public source polling does not change any browser or MCP calculation contract.

Optional, off-by-default fixed-category usage measurement writes only the isolated web3_studio_events table. /api/growth excludes QA unless ?qa=1; reports count opt-in events, not people, linked sessions, verified backlinks or revenue. A separate MEASURE_LIMIT binding caps submissions. /api/stats feedback is separate. Run node scripts/market-smoke.mjs after deploying and node scripts/growth-audit.mjs --check for source/discovery/measurement health. The existing daily fleet-autopilot retains 30 daily diagnostic records.

Research board (1.4)

/market.html now combines eight curated assets, daily history comparisons, source-linked research questions and a local evidence notebook. /api/research is a separate read-only feed; the original four-asset /api/market contract and deterministic MCP tools remain available. status describes snapshot retrieval; every quote has its own freshness status. Changes use daily samples within two hours of the intended interval, never synthetic candles. Missing history does not erase usable current prices.

Watchlist baselines and notes are written to browser storage only on explicit Save, expire after 30 days and have separate Delete actions. They are never uploaded. Export is a user-authored Markdown file, not automatic publication or an AI verdict. Four optional fixed-category events support evaluating the experiment without collecting notes, saved prices or asset selections. scripts/research-smoke.mjs verifies actual source/history connectivity after deployment; the existing daily audit covers this feed too.

Read-only aggregate coverage (2026-10-11)

GET and HEAD /api/growth and /api/stats execute SELECT statements only. Schema setup and 35-day retention cleanup happen on accepted event/feedback submissions, never during reporting. Missing storage or tables return HTTP 503 with status: "unavailable" and groups, daily, today set to null. An empty array is only returned after a successful query. Internal errors and raw inputs are not included in public responses.

Both endpoints include generatedAt, asOf, timezone: "UTC", scope, window, and metric definitions. groups retains its legacy dimensions: 28 rolling days for growth; 35 rolling days of feedback for stats. daily contains sparse date groups for the previous 28 complete UTC days; today is separate and partial. Every window states inclusive start and exclusive end. Every row includes site; the hub queries all sites without duplicating totals onto each tool host. A tool hostname returns only its own rows. Growth's qa=1 is QA-only and uncached, while default growth excludes QA. Feedback retains the qa dimension in all groups: business reports must filter qa=0 explicitly. Self-reported own_completed does not establish an independently verified task. Counts are events/submissions, not users, MCP demand, registrations or revenue.

Successful public reads advertise Cache-Control: public, max-age=3600; unavailable responses and growth QA reads use no-store. generatedAt records query generation, not a claim that source data is fully processed. This change adds no new event collection, D1 tables, persistent visitor IDs or scheduled job. For a deployment that must avoid synthetic business activity, set FLEET_READ_ONLY_PROBE=1: smoke scripts verify assets, metadata and sources but skip feedback/event submissions, aggregate reads, and MCP tools/call requests. The reporting collector can then fetch each aggregate exactly once.