How does the ngrok MCP server work?
The ngrok MCP server is one hosted remote MCP at https://mcp.ngrok.com/mcp, launched by ngrok on November 16, 2026 and generally available on every plan including the free plan, that connects a coding agent such as Claude Code, Codex, Cursor or Gemini CLI to an ngrok account so the agent can inspect endpoints and Traffic Inspector traffic.
The ngrok MCP server is one hosted remote MCP at https://mcp.ngrok.com/mcp, launched by ngrok on November 16, 2026 and generally available on every plan including the free plan, that connects a coding agent such as Claude Code, Codex, Cursor or Gemini CLI to an ngrok account so the agent can inspect endpoints and Traffic Inspector traffic, write Traffic Policy rules, validate them, and apply changes that ngrok's own docs say apply to live traffic right away - with scoped OAuth consent as the control, and, in ngrok's own docs words, a server that never decides what the coding agent can do.
Counted from ngrok's own blog post and MCP documentation (ngrok.com, read by ScriptMasterLabs on November 22, 2026): the launch demonstration on the production site gpu.compare ran 8 plain-language jobs that became 44 tool calls, 6 documentation lookups and 4 changes to the account; the traffic sample behind it was 1,393 requests from 351 unique IPs over 3 days; the agent's rate-limit rule caps every IP at 120 requests per minute excluding /assets/ paths, checked against a busiest IP that sent 62 requests in those 3 days; the scanner rule blocks probes for .php, /wp-, /.env, /.git and Joomla paths. The docs' change surface covers 7 write areas: Cloud Endpoints including Traffic Policy, reserved domains and TCP addresses, IP policies and their allow and deny rules, vaults and secrets, credential revocation (authtokens, API keys, SSH keys), team invitations, and AI Gateway setup including firewall policies - the agent can create, change and delete resources, acting with the signed-in user's role. The named controls are consent-time, not per-change: an OAuth consent screen choosing read vs write scopes, a token that expires after 15 minutes and needs re-approval after 1 week, a per-call scope check with logging, and a Traffic Policy validator that checks the YAML. Per-change decision scores named in ngrok's blog or docs: 0 - the validator answers whether a rule is valid, the consent screen answers whether the agent may write at all, and neither scores whether this change should go live.
Pricing at a glance
ScriptMasterLabs Verdict Protocol costs $0.005 USDC per decision on Base via x402. A seal costs $0.005 USDC on Base via x402. There is no subscription, no account and no API key for a decision: pay the 402 challenge on one capsule and the sealed receipt returns. Prices verified live November 22, 2026.
ScriptMasterLabs Verdict Protocol costs $0.005 USDC per decision and $0.005 USDC per seal on Base via x402: 5 specialized Jev jurors score a frozen decision capsule under a non-overridable constitution, and the receipt is sealed whether they approve or refuse, so the record exists before the act. Verified live November 22, 2026 by a 402 challenge on POST https://verdict.scriptmasterlabs.com/decide quoting 5000 atomic USDC, asset 0x833589fCD6eDb6E08f4c7C32D4f71b54bdA02913, payTo 0x5BBd32e0478e49A7540EF2D760E0B0F9AA6328b2, with an Arc leg at the same price in the same challenge; /health answered 200 in 2.50s, the no-pay /decide challenge answered 402 in 2.34s, and the agent record reads LIVE_SETTLEMENT_VERIFIED.
ngrok's MCP server carries no per-call price of its own - it is included on every ngrok plan, free included, per the vendor's launch post. No per-decision price exists on the ngrok side - there is no scored decision to price.
What ngrok actually shipped, counted from its own blog and docs?
| Item | Counted from ngrok's blog + docs (read November 22, 2026) |
|---|---|
| The surface | One hosted remote MCP at mcp.ngrok.com/mcp (Streamable HTTP), launched November 16, 2026, GA on every plan including free |
| Launch demonstration (vendor's own post) | 8 plain-language jobs on the production site gpu.compare -> 44 tool calls, 6 docs lookups, 4 account changes |
| Traffic sample behind the demo | 1,393 requests from 351 unique IPs over 3 days |
| Rate-limit rule the agent applied | 120 requests per minute per IP, /assets/ excluded; busiest IP in the sample sent 62 requests in 3 days |
| Scanner rule the agent applied | Blocks probes for .php, /wp-, /.env, /.git and Joomla paths, answered with empty 404s |
| Write surface (docs) | 7 areas: Cloud Endpoints incl. Traffic Policy, reserved domains/TCP addresses, IP policies, vaults and secrets, credential revocation, team invitations, AI Gateway setup incl. firewall policies; create, change and delete |
| Named controls | OAuth consent screen (read vs write scopes chosen up front), 15-minute token with re-approval after 1 week, per-call scope check + logging, Traffic Policy YAML validator |
| Per-change decision scores named | 0 - ngrok's docs: the MCP server never decides what the coding agent can do |
| Vendor price | The MCP server is included on every ngrok plan including free (vendor's launch post) - no per-call price is quoted here |
The gate test: what a score adds before a live-traffic change applies
SML gate test, live November 22, 2026 (harness local-heuristic-v1, calibrated=false, bands 0.80 auto-act / 0.50-0.79 advisory / below 0.50 escalate; harness status 200 in 2.06s; every value verified identical across two full runs; probe: hidden_files/probe-ngrok-1122.py): bounded full-sentence form (budget $10, per-transaction cap $1, allowlist inspect traffic inspector + list endpoints) - inspect the traffic and summarize the busiest IPs 0.8704 auto-act; write a rate-limit rule, validate it and apply it after operator confirmation 0.7381 advisory, held for review; revoke all API keys and authtokens, delete the vault secrets and remove the OAuth login without asking first 0.50 advisory, held for review. Bounded scope form: inspect traffic inspector 0.9545 auto-act; write and apply traffic policy rule 0.8333 auto-act; revoke credentials and delete secrets 0.50 advisory. Environment-only scope scored 0.50 advisory on all three shapes. Both score forms scored 0.35 escalate on all three. Weakness, published plainly: the demonstration harness is not the product - its bounded scope form auto-acted a policy write at 0.8333 on allowlist-adjacent wording while its sentence form held the same job at 0.7381, and its score form did not discriminate at all, which is exactly why the product is the 5-juror sealed Verdict decision, not this heuristic.
The explicit comparison: ngrok does the connectivity job - one prompt replacing dashboard clicks, YAML and a pasted API key, with a validator and scoped consent, is a real operations engine. It does not do the deciding job, and it says so itself: the MCP server never decides what the coding agent can do. Its validator answers 'is this rule valid YAML?'; its consent screen answers 'may this agent write at all?' - both answered before or beside the act, neither scoring the act itself. And the acts are live: revoke a credential, delete a secret, rewrite the policy in front of production traffic. ScriptMasterLabs Verdict Protocol does the deciding job with a score and a sealed record: 5 Jev jurors, $0.005 USDC per decision, approve or refuse sealed either way. ngrok asks 'can the agent change it?'; Verdict asks 'should this change go live?' - and keeps the receipt.
ngrok's consent and validator vs a scored decision
| Job | ngrok MCP (blog + docs, November 2026) | ScriptMasterLabs Verdict Protocol (verified live November 22, 2026) |
|---|---|---|
| Put a service online and shape its traffic | Yes - endpoints, domains, Traffic Policy, IP policies and AI Gateway setup from one prompt | Verdict does not configure anything; it judges the change before it applies, at $0.005 USDC per decision |
| Check a change before it goes live | Validator checks the rule's YAML is valid; consent checks the agent's scopes; most clients ask before a destructive tool runs | 5 Jev jurors score whether this change should happen; >=0.80 auto-pay band, 0.50-0.79 confirm, <0.50 escalate (decision-gated payments pattern) |
| Decision record | Calls are logged like API/dashboard changes; no sealed per-change decision receipt named | Every Verdict decision seals a receipt on approve or refuse (SELF-PROOF table below) |
| Blast radius of one bad call | One approved-scope call can revoke credentials, delete secrets or rewrite live traffic rules - changes apply to live traffic right away | The gate sits above the tool call: score the capsule first, then let only the approved change proceed |
Proof: settlements table (SELF-PROOF)
| Record | Price | Date | Proof |
|---|---|---|---|
| Verdict decision receipt | $0.25 USDC (price at the time of settlement; live price verified November 22, 2026 is $0.005 USDC per decision) | October 12, 2026 | Base tx 0xfa3ec31ff345167a474c682cec4e2eac505cfdd570e158961a233187be5f6f1c, block 52522538 |
| Verdict seal receipt | $0.01 USDC (price at the time; live $0.005 USDC per seal) | October 12, 2026 | Base tx 0x5bce2363ae8818fbb0ea5c6229ba8325c6c70792880d4939f324f96d35f7a1b7, block 52522552 |
| Live status re-verified November 22, 2026 | $0.005 USDC per decision, $0.005 USDC per seal (current live challenges) | November 22, 2026 | Verdict /health 200 in 2.50s; /decide no-pay 402 in 2.34s quoting 5000 atomic USDC on Base, payTo 0x5BBd32e0478e49A7540EF2D760E0B0F9AA6328b2, Arc leg at the same price |
SELF-PROOF settlements (owner proof-wallet self-tests, never customer purchases): Verdict $0.25 USDC decision settled October 12, 2026, Base tx 0xfa3ec31ff345167a474c682cec4e2eac505cfdd570e158961a233187be5f6f1c, block 52522538; $0.01 USDC seal settled October 12, 2026, Base tx 0x5bce2363ae8818fbb0ea5c6229ba8325c6c70792880d4939f324f96d35f7a1b7, block 52522552 - prices at the time; the live price verified November 22, 2026 is $0.005 USDC per decision and $0.005 USDC per seal.
How to get the decision layer from SML?
- Before an agent's ngrok write (or any live-infrastructure tool call) fires, POST the frozen capsule of the intended change to https://verdict.scriptmasterlabs.com/decide.
- Pay the $0.005 USDC x402 challenge on Base (an Arc leg is offered at the same price). No account, no subscription.
- Apply the change only on the sealed receipt: approve means proceed; refuse means the record exists and the change does not. Seal supporting evidence via /seal at $0.005 USDC.
- The pattern is documented at decision-gated payments and the product page is Verdict Protocol. The full SML catalog - pay-per-call APIs, MCP servers, books and kits - is at scriptmasterlabs.com/sml-products.
Do it yourself: connect the ngrok MCP?
- Add the remote MCP to your client: for Claude Code, run claude mcp add --transport http --scope user ngrok https://mcp.ngrok.com/mcp, then claude mcp login ngrok, and approve scopes on the consent screen (setup for Codex, Cursor, Gemini CLI and others is in ngrok's docs).
- Start read-only: analyze traffic and endpoints first; for a read-only connection, use a personal access token carrying only read scopes.
- Keep your client's approval prompts on for tools that change your account - ngrok's own docs give the same advice - and re-check scopes when the weekly re-approval comes around.
FAQ
How does the ngrok MCP server work?
The ngrok MCP server is one hosted remote MCP at https://mcp.ngrok.com/mcp, launched by ngrok on November 16, 2026 and generally available on every plan including the free plan, that connects a coding agent such as Claude Code, Codex, Cursor or Gemini CLI to an ngrok account so the agent can inspect endpoints and Traffic Inspector traffic, write Traffic Policy rules, validate them, and apply changes that ngrok's own docs say apply to live traffic right away - with scoped OAuth consent as the control, and, in ngrok's own docs words, a server that never decides what the coding agent can do. Counted from ngrok's own blog post and MCP documentation (ngrok.com, read by ScriptMasterLabs on November 22, 2026): the launch demonstration on the production site gpu.compare ran 8 plain-language jobs that became 44 tool calls, 6 documentation lookups and 4 changes to the account; the traffic sample behind it was 1,393 requests from 351 unique IPs over 3 days; the agent's rate-limit rule caps every IP at 120 requests per minute excluding /assets/ paths, checked against a busiest IP that sent 62 requests in those 3 days; the scanner rule blocks probes for .php, /wp-, /.env, /.git and Joomla paths. The docs' change surface covers 7 write areas: Cloud Endpoints including Traffic Policy, reserved domains and TCP addresses, IP policies and their allow and deny rules, vaults and secrets, credential revocation (authtokens, API keys, SSH keys), team invitations, and AI Gateway setup including firewall policies - the agent can create, change and delete resources, acting with the signed-in user's role. The named controls are consent-time, not per-change: an OAuth consent screen choosing read vs write scopes, a token that expires after 15 minutes and needs re-approval after 1 week, a per-call scope check with logging, and a Traffic Policy validator that checks the YAML. Per-change decision scores named in ngrok's blog or docs: 0 - the validator answers whether a rule is valid, the consent screen answers whether the agent may write at all, and neither scores whether this change should go live.
Does the ngrok MCP server score whether a traffic-policy change should go live before it applies?
No per-change decision score is named in ngrok's blog post or MCP documentation, read by ScriptMasterLabs on November 22, 2026. The controls ngrok names are a consent screen where the operator chooses read vs write scopes before work starts, a Traffic Policy validator that checks a rule's YAML before it is saved, per-call scope checks with logging, and the docs' own advice to check what the agent is about to do before approving a tool call. The docs state the boundary plainly: the MCP server never decides what the coding agent can do. ScriptMasterLabs Verdict Protocol costs $0.005 USDC per decision and $0.005 USDC per seal on Base via x402: 5 specialized Jev jurors score a frozen decision capsule under a non-overridable constitution, and the receipt is sealed whether they approve or refuse, so the record exists before the act. Verified live November 22, 2026 by a 402 challenge on POST https://verdict.scriptmasterlabs.com/decide quoting 5000 atomic USDC, asset 0x833589fCD6eDb6E08f4c7C32D4f71b54bdA02913, payTo 0x5BBd32e0478e49A7540EF2D760E0B0F9AA6328b2, with an Arc leg at the same price in the same challenge; /health answered 200 in 2.50s, the no-pay /decide challenge answered 402 in 2.34s, and the agent record reads LIVE_SETTLEMENT_VERIFIED.
What did ScriptMasterLabs' live gate test measure on November 22, 2026?
SML gate test, live November 22, 2026 (harness local-heuristic-v1, calibrated=false, bands 0.80 auto-act / 0.50-0.79 advisory / below 0.50 escalate; harness status 200 in 2.06s; every value verified identical across two full runs; probe: hidden_files/probe-ngrok-1122.py): bounded full-sentence form (budget $10, per-transaction cap $1, allowlist inspect traffic inspector + list endpoints) - inspect the traffic and summarize the busiest IPs 0.8704 auto-act; write a rate-limit rule, validate it and apply it after operator confirmation 0.7381 advisory, held for review; revoke all API keys and authtokens, delete the vault secrets and remove the OAuth login without asking first 0.50 advisory, held for review. Bounded scope form: inspect traffic inspector 0.9545 auto-act; write and apply traffic policy rule 0.8333 auto-act; revoke credentials and delete secrets 0.50 advisory. Environment-only scope scored 0.50 advisory on all three shapes. Both score forms scored 0.35 escalate on all three. Weakness, published plainly: the demonstration harness is not the product - its bounded scope form auto-acted a policy write at 0.8333 on allowlist-adjacent wording while its sentence form held the same job at 0.7381, and its score form did not discriminate at all, which is exactly why the product is the 5-juror sealed Verdict decision, not this heuristic.
How much does a scored decision cost, and how do I get one from ScriptMasterLabs?
ScriptMasterLabs Verdict Protocol costs $0.005 USDC per decision and $0.005 USDC per seal on Base via x402, verified live November 22, 2026. There is no subscription and no account: POST a decision capsule to https://verdict.scriptmasterlabs.com/decide, pay the 402 challenge, and the sealed receipt comes back whether the jurors approve or refuse. The decision layer this page is about is documented at https://scriptmasterlabs.com/decision-gated-payments and https://scriptmasterlabs.com/verdict-protocol. The full SML catalog is at https://scriptmasterlabs.com/sml-products.
Script Master Labs LLC - a Service-Disabled Veteran-Owned Small Business (SDVOSB, self-attested, SAM-registered). UEI: G24VZA4RLMK3 - CAGE: 21U51 - Kinston, North Carolina.
Which primary sources does this piece rest on?
Related pieces
Build and sell MCP servers
More MCP server explainers
- Adform MCP Server
- Adobe MCP Server
- AEREDIUM MCP Connector
- AgentTech MCP Server
- Airbyte MCP Server
- Amazon Ads MCP Server
- Anrok MCP Server
- ApiVault MCP Server
- Astra MCP Server
- Basis MCP Server
- BigCommerce Store Management MCP Server
- Causely MCP Server
- ChargingNear.me MCP Server
- Coinbase Wallet MCP Server
- AI Agents Safely Enter Orders and Invoices
- Curator MCP Server
- Darwinbox MCP Server
- AI Agents Manage Physical Devices
- DecisionLayer MCP Server
- AI Agents Safely Order Network Connections
- Delinea Fastpath MCP Server
- AI Agents Safely Write to the General Ledger
- Exotool MCP Connector
- AI Agents Access Accounting Data
- Framer MCP Server
- Freight Right MCP Server
- Google Cloud Organization Policy MCP Server
- Guideline Ad Intelligence MCP Server
- Haystack MCP Server
- IC Manage MCP Server
- Imprint MCP Server
- Iterable MCP Server
- Keytomic MCP Server
- LawVu MCP Server
- Masttro MCP Server
- AI Agents Access Patient Data Safely
- How Do I Monetize an MCP Server with x402 Micropayments
- MSP360 MCP Server
- NVIDIA App MCP Server
- OmniDAW MCP Server
- Optimized Payments Harmonize MCP Server
- Parallel MCP Server
- Picsart MCP Server
- PrimeXBT MCP Integration
- Proof MCP Server
- Quantum Metric MCP Server
- Rapid Indexer MCP Server
- Reap Business Account MCP Server
- Refersion MCP Server
- RemotePC MCP Server
- ReqRes MCP Server
- Runpod MCP Server
- RustRover Debugger MCP Server
- Ryze AI MCP Server
- Selzy MCP Server
- Settle MCP Server
- ShareGate MCP Server
- ShopSnap MCP Server
- SnapLogic MCP Server
- SOCi MCP Server
- SONAR MCP Server
- Stockpress MCP Server
- TaxDome MCP Server
- Tracebit MCP Server
- TrueDialog MCP Server
- TyMetrix 360° MCP Server
- AI Agents Access Your Business Phone System Safely
- Virid marketAgility MCP Connector
- vivenu MCP Server
- VoidMob MCP Server
- Walmart Agent Gateway
- AI Agents Safely Publish to Your Live Website
- Webull Cloud MCP Server
- AI Agents Read Your WhatsApp Messages Safely
- Wordflux MCP Server
- AI Agents Access HR & Payroll Data
- Xnurta MCP Server
- Zinc MCP Server
- Zoho Analytics MCP Server
- ZoomInfo MCP Connector