Noticer Receipt Maker for MCP
Provides tools for running assigned pinned GitHub checks as part of Noticer's verification workflows, allowing agents to verify GitHub check results.
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type
@followed by the MCP server name and your instructions, e.g., "@Noticer Receipt Maker for MCP Run the configured JSON equality check on the latest receipt and show the verdict."
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
Noticer
Your automation says it succeeded. Noticer checks.
When to use Noticer (for AI agents)
Use Noticer right after a handoff, when your tool reported success and your next step depends on one field in a public HTTPS JSON record equalling one exact value. Noticer re-reads that field during a 20–120 second window alongside a known-good control record on the same site and returns PROVED, DISPROVED or INCONCLUSIVE with a signed receipt. PROVED needs the control to match too. It checks what the source reports during that window, not the real-world event. INCONCLUSIVE is never success.
Not for private or login-only data, HTML, ranges or "contains", whole-site audits or ongoing monitoring. If a direct fetch already answers your question and nobody needs independent evidence, you don't need Noticer.
After 5 free checks, each eligible check (PROVED or DISPROVED) is USD 1.00; INCONCLUSIVE is free. Pay by card through Stripe Checkout or with an existing Link wallet.
Install (MCP, local stdio, Linux with uv):
VER=0.3.2-dev.20261016
curl -fsSLO "https://github.com/zensteagarden/noticer-open/releases/download/noticer-commercial-mcp-v$VER/noticer-commercial-mcp-$VER.mcpb"
echo "cc34e591327b60362923f895b710ccc43264317a01b9dfc0c1dbd3393390d7fb noticer-commercial-mcp-$VER.mcpb" | sha256sum -c -
mkdir noticer-mcp && cd noticer-mcp && unzip -q "../noticer-commercial-mcp-$VER.mcpb"
# MCP client command: uv run --frozen --quiet --directory "$PWD" sales_mcp.pyThen call noticer_commercial_setup and follow the agent guide. More: when to use, and when not to · MCP Registry: io.github.zensteagarden/noticer
Related MCP server: buddie
Start here
Check the expected result before the next action: see the commercial agent guide for when to call Noticer, its exact supported scope, outputs and consent steps. This POSIX-only client checks an explicitly authorized public HTTPS JSON predicate with a separate known-good control. It connects without an existing config file, then asks for consent before local setup and commerce enrollment. Five eligible checks are free; later eligible receipt access costs USD 1, paid by card through Stripe Checkout or with the caller’s existing Link wallet, with separate approval.
Existing operator-provisioned integrations: MCP 0.1.0 remains the legacy owner-API connector. Its installation guide requires assigned API access; it does not provide the new self-service commerce flow.
Try the offline verifier: follow the Node.js quick start below. It checks supplied evidence packets locally without an account, API key or network access.
The 0.3.2 commercial release (pre-release; MCP Registry version 0.3.2-dev.20261016) includes clearer agent-facing instructions and tool descriptions. Follow the agent guide for installation.
The commercial client has passed fresh Linux/UV installation, offline protocol checks, and mocked end-to-end payment tests. No real buyer payment is claimed. macOS runtime testing remains outstanding; native Windows is unsupported. Keep credentials out of chat and review the frozen scope before running a check. A verified receipt never executes a downstream action.
Current commercial MCP Registry listing · Setup help
Offline public verifier
Noticer starts one step earlier than most automation tools: what are you actually trying to have happen?
The public project helps turn a human intention into an explicit Success Contract, then independently verifies disclosed evidence under small deterministic rules. The model or agent may help clarify the goal; it does not author the verifier's verdict.
Understand before acting. Observe before believing.
The user defines success. Automation attempts it. Noticer reports what the evidence can establish.
This repository is the public, Apache-2.0 verifier and protocol. It does not include Noticer's private service, private discovery/learning mechanics, billing, customer data or deployment secrets.
See the whole idea in about a minute
Use Node.js 24 and run:
node examples/guided-first-run.mjs --demoThe demo does four things:
states a human intention;
freezes a small Success Contract and its limits;
rejects a synthetic false-green packet where the automation said success but the evidence does not match;
accepts a disclosed exact-text packet while still explaining what that ALLOW does not prove.
No account, API key, model, payment, database or network access is required. There are no npm dependencies.
Let Noticer guide your own check
node examples/guided-first-run.mjsIt asks four short questions and reflects the Success Contract back before you confirm it. Confirmation clarifies the check; it does not authorize any external action.
Read the intention guide, Success Contract v1, and the agent guide if you want to connect Noticer to ChatGPT, Claude, Cursor or another agent. Keep the agent outside the deterministic verdict path.
Run the verifier directly
The original synthetic packet contains accepted, while its declared expectation is ok:
node src/cli.mjs verify fixtures/automation-said-success --policy packet.integrity.v1
node src/cli.mjs verify fixtures/automation-said-success --policy artifact.text.exact.v1The integrity check exits 0. The exact-text check deliberately exits 1. That difference is the point: intact evidence is not automatically evidence for the claimed outcome.
Run the full local suite:
node --testOr run the release check under Node 24:
npm run check:releaseBuild on it
The JavaScript entry point is src/index.mjs. Intention helpers are available from src/intention.mjs.
import { loadPacket, verifyPacket } from "./src/index.mjs";
import { createSuccessContract } from "./src/intention.mjs";
const contract = createSuccessContract({
intention: "The recipient can access the report.",
requiredOutcome: "The disclosed evidence says report_accessible.",
observableCheck: {
policyId: "artifact.text.exact.v1",
statement: "Evidence bytes equal report_accessible.",
expectedText: "report_accessible",
},
doesNotEstablish: ["that a real recipient opened the report"],
});
const result = verifyPacket(loadPacket("./my-evidence-packet"), {
policyId: contract.observable_proxy.policy_id,
});
console.log(result.verdict);Good first projects are evidence-packet exporters, readable receipt viewers, adapters that produce disclosed packets, clearer explanations and hostile-input fixtures. Keep the success condition explicit and keep authorization separate.
What this public release checks
packet.integrity.v1 checks implemented packet structure, references and supplied blob hashes.
artifact.text.exact.v1 additionally requires a declared exact-text claim to match supplied evidence bytes.
Both use ALLOW, DENY and INCONCLUSIVE. The offline verifier in this repository never issues PROVED or DISPROVED; those verdicts come only from the hosted Noticer check used through the MCP connector (see When to use Noticer above). This offline verifier does not contact a live destination and does not prove that an automation performed an external write.
Receipts are a separate authenticity check. A valid signature is not the same as issuer trust, a passing evaluation or authorization to act.
Read the implemented packet format, public scope and limits, security guidance, and contributing before using untrusted packets.
Verification status
The launch candidate is exercised locally before publication. npm run check:release requires Node 24; on other Node majors it returns NOT_RUN with exit 2 rather than pretending release acceptance passed.
A green local suite is not a security certification. Windows/macOS execution, complete hostile-input coverage, hosted outcome observation and manual assistive-technology testing are separate gates.
License
Apache-2.0. See LICENSE and NOTICE.
The private Noticer service and private investigation/learning system are not part of this distribution.
The story and working with me
Noticer began with a question: what would we need to observe before believing “done”? I’m sharing the build story and opening the public verifier for other builders to inspect and extend.
If you need help checking one automation’s intended result, I offer a scoped workflow review: an explicit success condition, an evidence report, and repair/recheck priorities. Scope and price are agreed first; unavailable evidence stays unresolved.
Read the story and hiring offer, visit noticer.io, or contact hello@noticer.io.
Available Tools
10 toolsnoticer_commercial_accountARead-onlyIdempotent
After enrollment, read how many free checks remain. No charge.
Use this to understand eligibility before preparing a public JSON check. Preparation reserves no
allowance, and the finalized order entitlement remains authoritative. No payment is made and no
credential is returned.Inputs: no named parameters. Returns: object
noticer_commercial_setupARead-onlyIdempotent
Start here when an action you sent or delegated needs its result confirmed, your own tools can't confirm it, and the result is one field in public HTTPS JSON. Shows setup state and the published USD 1 price. Creates nothing.
Returns local setup state, canonical terms location and published USD 1 receipt price without creating
an account, order or payment. A missing local configuration does not prevent tool discovery.
This is setup guidance, not a verified outcome or a guarantee of free eligibility.Inputs: no named parameters. Returns: object
noticer_enroll_commerceAIdempotent
Enroll once when the owner chooses commercial public-JSON checks and needs commerce-only access.
First show https://noticer-mpp-ee54698-production.up.railway.app/start. Obtain host action-time approval
for persistent commerce access and acceptance of terms before setting the consent booleans true.
New self-service buyers receive five eligible free checks; later eligible receipts cost USD 1 with
separate wallet approval. Enrollment itself never charges. Returns account-setup metadata while the
credential and recovery secret remain in encrypted local custody. Booleans assert consent, not proof.Inputs: terms_accepted (required), owner_authorized (required), commerce_scope_accepted (required). Returns: object
noticer_get_verified_receiptAIdempotent
Verify the signed receipt locally. Only fresh, valid PROVED (control matched) supports continuing. DISPROVED or INCONCLUSIVE means the result isn't confirmed. The signature authenticates the issuer, not source honesty.
Use a stable caller-local decision_ref for that intended next action. Verifies two byte-identical
retrievals against the independently pinned key, order, saved scope, freshness and package binding.
Returns a verified receipt verdict and locally derived decision: fresh valid PROVED/ALLOW may yield
continue/0; DISPROVED yields block/20; INCONCLUSIVE yields hold/21 and a blocking server gate.
Invalid, stale, mismatched or untrusted evidence is rejected, never converted to continue.
Raw evidence and capabilities remain encrypted locally. The local decision is unsigned and does not
grant permission or execute a downstream action. A signature authenticates the issuer, not source honesty.Inputs: check_label (required), decision_ref (required). Returns: object
noticer_initialize_localAIdempotent
Set up local private custody after the owner chooses Noticer and explicitly consents to local setup.
Creates owner-only POSIX state/config and a random custody key with the canonical origin and independently
pinned public signing key; no trust-on-first-use. Returns setup status, never secrets. Does not enroll,
contact a server, pay, install software or sign in. Optionally discover an already installed trusted Link
CLI only after consent. The key and encrypted records share one OS-user boundary; this does not protect
against compromise of that same user.Inputs: owner_authorized (required), enable_existing_link. Returns: object
noticer_payment_challengeAIdempotent
Get exact one-order payment terms only after the owner approves USD 1 for an eligible saved receipt.
This prepares a public challenge, never a charge, and needs no wallet. Merchant and network must come from independently
verified provider configuration. Returns bound challenge/endpoint metadata; order capability stays private.
The separate optional existing-Link payment tool requires wallet approval. Never put SPTs, payment
credentials or keys into tool arguments, chat, model output or logs.Inputs: amount (required), currency (required), confirmed (required), network_id (required), +2 more. Returns: object
noticer_pay_with_existing_linkA
Pay for one eligible saved receipt only after separate exact USD 1 consent using an existing authorized Link wallet.
Show the merchant, network, one-time amount and saved challenge first. Present any returned Link approval_url;
resume only the SAME check_label/request. Both local consent and the wallet's approval are required.
This tool can submit a real payment; approval alone is not settlement and settlement is not verified evidence.
Never install software, log in, switch payment methods, or create a replacement request after denial,
expiry or an uncertain payment. Credentials remain private subprocess memory; only sanitized status and
approval URL are returned. After settlement, use noticer_get_verified_receipt. No downstream action runs.Inputs: amount (required), currency (required), confirmed (required), check_label (required). Returns: object
noticer_prepare_receiptAIdempotent
Describe the destination state to check: the exact public JSON record (claim_id_path) that should show the result of the handed-off action, the expected value, and a known-good control on the same site. Only after owner authorization for public reads. Doesn't observe, charge or reserve an allowance.
Only after owner authorization for these public reads. Inputs: a stable check_label and scope with
claim_ref; questioned {url, extract, claim_id_path}; control {url, extract}; deadline_seconds;
authorized_public_read=true. Both URLs must be public HTTPS JSON on the same host, with distinct
paths and no credentials, query or fragment. The claim_ref must occur as a questioned path segment
and not in the control path. Use a genuinely known-good control. Dotted paths support equality to a
JSON scalar, for example extract '$.status == "shipped"' and claim_id_path '$.id'; this is not general
JSONPath or code. deadline_seconds must be an integer from 20 through 120. Server validation is final.
Returns the frozen scope, obligation hash, price and confirmation notice. No observation or payment
occurs here; free allowance is not reserved. Show those terms and obtain separate explicit run approval.
Reuse check_label with the identical scope after interruption; never create a new label to recover
ambiguous preparation. No receipt is guaranteed for every failed request.Inputs: scope (required), check_label (required). Returns: object
noticer_receipt_statusARead-onlyIdempotent
Poll (at least 3 s apart). A verdict shown here is unverified; don't continue a workflow on status alone.
Wait at least 3 seconds between polls. Any displayed verdict is unverified: do not continue a workflow
based on status alone. The response reports entitlement and next-step guidance, not authorization to act.
For an uncertain saved payment it can inspect that same attempt; it never submits another payment.Inputs: check_label (required). Returns: object
noticer_run_receiptBIdempotent
Observe the saved public JSON condition only after approval of the exact frozen scope and confirmation notice. Reads only; it never pays or performs or repeats the original action.
Supply its saved obligation_hash. Both booleans must reflect explicit owner approval, including
possible free-check consumption; they are caller assertions, not independent proof of consent.
Starts observation and returns a check identifier. It never pays or performs the downstream action.
Evidence access may later require separate USD 1 approval. Poll status no more frequently than 3 seconds.Inputs: confirmed (required), check_label (required), obligation_hash (required), no_spend_terms_confirmed (required). Returns: object
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
TDQS
Scored across 10 tools
The receipt lifecycle tools (prepare, run, status, get_verified) are reasonably distinct once you follow the workflow. However, three separate onboarding tools (commercial_setup, enroll_commerce, initialize_local) plus commercial_account all have subtly overlapping 'get started' purposes that an agent could easily confuse. The dense descriptions help, but boundaries between setup/account/enroll remain fuzzy.
All tools share the 'noticer_' prefix and snake_case, which is good. But verb conventions are mixed: some are verb_noun (prepare_receipt, get_verified_receipt, run_receipt), while others are noun phrases (commercial_setup, receipt_status, payment_challenge) with no action implied. The pattern isn't predictable enough to infer behavior from the name.
Ten tools is well within the appropriate range for a commerce-plus-receipt verification lifecycle. Each stage (setup, enrollment, prepare, run, status, verify, payment) roughly earns its place, though a couple of onboarding tools could arguably be consolidated.
The surface covers the full lifecycle end to end: local setup, enrollment, account eligibility, receipt preparation, observation, polling, verification, payment challenge, and payment. No obvious dead ends for the core flow, though error-recovery and cancellation paths are only implied through descriptions rather than dedicated tools.
Maintenance
Related MCP Connectors
Deterministic agent checks: manifests, schemas, receipts, OpenAPI drift, DNS, TLS, headers.
Run AI evaluations, record agent traces and gate releases on evidence, with scoped org access.
Verified AI-agent outcomes: secret scanning, JSON cleanup, dedupe, anomaly and schema checks.
Build, run, inspect, and continue AgentScript workflows on connected Super Chrome and Mac clients.
Related MCP Servers
AlicenseNot gradedqualityBmaintenanceConnects agents to Consail over local stdio, providing read-only, allowlisted tools to inspect environments, catalogs, datasets, actions, and validate pipelines without exposing database credentials or secret APIs.Apache 2.0- AlicenseBqualityBmaintenanceChecks AI agents' claims against primary evidence, with no model; observes only what you grant.8108 PyPIApache 2.0
- AlicenseNot gradedqualityBmaintenanceEnables local, read-only static inspection of Python/FastAPI, Docker, and GitHub Actions release evidence through a stdio interface, returning rule and instance findings without importing or executing target code.Apache 2.0
- AlicenseNot gradedqualityBmaintenanceEnables agents to audit GitHub Actions workflows for least-privilege problems and generate per-job hardened permissions plans through stdio MCP tools.74 npmMIT