MCP 2026-07-28 · OpenAPI 3.1 · OAuth 2.1

Let AI assistants workwith your business.

Knotbind builds the connection that lets ChatGPT, Claude and Grok answer questions, take bookings and look up orders in your systems. Then we watch it, so a change on either side never breaks it silently.

Check if agents can use your site Contact us See it on a shop

ChatGPT
Claude
Grok
Gemini
Any MCP client

Read the documentation · docs.knotbind.com

Are you an AI agent?Get onboarding in one fetch.

curl -sS https://www.knotbind.com/agents.md Read the quickstart →

Free to start. Pro is $199 per month. Business is $399 per month. Enterprise is agreed in writing.

Assistant● Nomad Outfits · via Knotbind

Do you have the Trailhead rain jacket in blue, size M?

→ check_stock({ sku: "RJ-204-BLU", size: "M", store: "lisbon-chiado" })
← 200 · 3 in stock · 84ms

Yes, 3 left at the Chiado store in Lisbon. Want me to hold one for you?

Please, I can come by after work.

→ hold_item({ sku: "RJ-204-BLU", size: "M", until: "tomorrow 20:00" })
← 201 · hold #H-3318

Done. It's held under your name until 20:00 tomorrow.

Illustrative example · fictional businesses

// built for agents

Knotbind is software agents can use, too.

An agent can sign up, connect an MCP server, run Checks and read Findings without a human. Approving a production change stays with a person.

REST API

Every dashboard action over JSON, with scoped API keys.

/openapi.json

MCP server

Add Knotbind itself in Claude, ChatGPT or Cursor and ask it about your integration.

https://www.knotbind.com/mcp

CLI

The same operations from a terminal or a script, JSON out.

knotbind readiness example.com

Webhooks

Receive signed HTTPS notifications when Checks fail, recover, or drift is detected.

HMAC-SHA256 signed with replay protection

// what knotbind is

The knot between AI assistants and the systems you already run.

An assistant can only help your customers if it can reach your stock, calendar or orders. Knotbind is that link: we build it, route it, and keep it working. Every call is monitored — SSRF guards block internal targets, secrets and PII are redacted from answers, and the kill switch plus Owner Approval gate stop any abuse before it lands (see /security).

customer

ai assistants

knotbind

customer system

answer

Your customer Asks a question in the assistant they already use.
The assistant ChatGPT, Claude, Grok or any assistant that supports MCP.
Knotbind Your MCP server and gateway: what the assistant may do, done safely, watched every hour.
Your system Shop, booking tool, database or spreadsheet. Built for you if you have no API yet.

// under the hood

Built on open standards, not screen scraping.

Every connection is a typed MCP server over your OpenAPI description, authenticated with OAuth 2.1 and checked on a schedule.

knotbind — christine-flowers

$ knotbind connect https://mcp.christine-flowers.example/mcp{ "integration": "acct_christine_flowers", "tools": 3 } $ knotbind status{ "health": "passing", "checks": 3, "every": "hourly" } $ knotbind run{ "passed": false, "create_booking": "422 ×2" } $ knotbind finding{ "drift": "location now required", "state": "diagnosed" }

knotbind — finding → resolve

$ knotbind finding{ "fix": "add location="lj"", "sandbox": "12/12 passed" } # owner approves in the dashboard$ knotbind runs --limit 1{ "passed": true, "gate": "approved", "state": "resolved" }

protocol: MCP 2026-07-28 over Streamable HTTP
api contracts: OpenAPI 3.0 and 3.1, breaking-change diff on every version
auth: OAuth 2.1 with scoped read, write and tool access
mcp registry: Several servers, per-tool contract checks, workflow checks and schema drift alerts
gateway: OpenAI-compatible endpoint, per-key budget and request rate. Cache and fallback are modules.
safety: Sandbox-only checks, SSRF guards, secrets redacted from traces
credential vault: Sealed upstream secrets, 15–60 min operation-bound grants
audit: Append-only record of every run, finding and approval

// services

What we offer.

Four services, then optional Gateway modules. Each one exists because a connection breaks, a bill runs away, or a reply leaks something it should not. See the problem and what we do about it, or follow one shop through it.

Build for small businesses

No API, no developers, no problem. We build your API and MCP server from scratch, then connect it to the AI assistants your customers already use — Claude, ChatGPT, Grok, Meta Muse, and many more.

API built or described for your existing system
OpenAPI to MCP: preview, export executable packages, or deploy a private endpoint
Owner-confirmed Sandbox Checks with explicit read-only fixtures
Encrypted operation credentials and temporary API-key/OAuth consent
One MCP server that works across all assistants
Tested and verified in up to three assistants
Complete source code delivered to you
Monitoring and health checks set up on handover

Included in Business

OpenAPI generator features and limits

How a Build works

Reliability

Automated Checks call every assistant action on a schedule, against every API and MCP server you connect. When something breaks, Knotbind diagnoses the issue, drafts a fix, and sends you a tested solution ready to approve.

Included in Pro and Business

See how it works

Knotbind Gateway

One OpenAI-compatible endpoint for the models your applications call. Each app gets an access key with a monthly budget and a request rate. Your provider keys stay in the workspace.

Per-key usage is included. Opt-in error hints name only the validation fields the provider already identified. Fallback, semantic caching, response redaction and workspace-wide analytics are separate modules.

Agent-Friendly Error Rewriting stops infinite retry loops that rack up API bills: when a provider returns a validation error, the Gateway rewrites it into a plain hint naming the exact field, so the assistant fixes the input instead of retrying the same call 20 times. For a typical workspace this shaves $800–$2,400 in wasted tokens per month.

Gateway modules · Error handling and Gateway limits · Business credential management

Included in Pro and Business

About the Gateway

Hosting regions

Gateway, monitoring, and all stored data kept in the region you pick — the European Union with a GDPR data processing agreement and EU AI Act compliance tooling, or the US under standard terms.

EU included in Business

About hosting regions

// gateway modules

Add a control only when you need it.

Semantic Cache, Content Guardrails, Model Failover, Advanced Rate Limits and Request Analytics are optional monthly add-ons on a Pro or Business Gateway. See module pricing.

// how reliability works

From a broken call to an approved fix, in six steps.

Every step is recorded. You are only asked once there is a fix that already works.

1 ⚙
Check

Scripted Checks call each action on a schedule, in your Sandbox.
2 ⚙
Detect

Two failures in a row, or a breaking change in the API description, open a Finding.
3 ⚙
Diagnose

Knotbind names the change behind the failure in plain words.
4 ✨
Propose

It drafts the TypeScript patch, re-runs the Check against it, and opens an automated Pull Request on your GitHub repository.
5 👤
Approve

Only a fix that passed reaches the Owner. Nothing ships without a yes.
6 ⚙
Resolve

The next passing Run closes the Finding.

// why connections break

Assistants fail quietly. Customers just leave.

See the scenario example.

Your system changes

A new required field, a renamed product, a moved endpoint. The assistant keeps calling the old way.

The assistant changes

Assistants update how they pick and call tools. A connection that worked last month can stop being used.

Nobody is watching

There is no error page in a chat. Without Checks you learn about it from a lost booking.

// who it is for

Businesses that want to be reachable by AI, and teams that already are.

Small businesses

Shops, salons, clinics, restaurants and service firms with no API yet. Follow a shop through it · See the Build service

AI SaaS companies

Your API and your MCP server change on different schedules. Knotbind tells you when they stop agreeing.

Automation agencies

Many client connections, one view of which ones are broken and why.

// how we work

Three rules we don't bend.

Nothing ships without a yes

Every fix needs the named Owner's approval. Fixes that fail testing never reach you.

Sandbox first

We test against a safe copy of your system, never against live customer data.

Everything is recorded

Each Check, finding and decision goes into an Audit Record that is never edited. Security details

// questions

Answered plainly.

What is Knotbind?

Knotbind connects businesses to AI assistants such as ChatGPT, Claude and Grok, and keeps those connections working. We build the API and MCP server an assistant needs, run a gateway between your apps and AI models, and watch every connection so a change on one side does not silently break the other.

What is an MCP server, in plain words?

MCP (Model Context Protocol) is the standard way AI assistants call outside tools. An MCP server is the small piece of software that tells an assistant what your business can do, such as check stock or book a slot, and carries out the request safely.

Does Knotbind change anything on its own?

No. Every fix is first tested in a Sandbox, a safe copy of your system, and only applied after the named Owner approves it. Each step is written to an Audit Record that is never edited.

How is this different from API monitoring?

Monitoring tells you an endpoint returned an error. Knotbind maps the failure to the assistant action it breaks, explains the change behind it, and hands you a fix that already passed in the Sandbox.

What happens when our MCP server changes?

Knotbind keeps a registry of your MCP servers and fingerprints every tool schema. A change that breaks callers — a removed tool, a new required parameter, a changed type — opens a Finding and reaches you by email, Slack or webhook. Each server also has its own health, so you can see which one is down and which one is fine.

Can Knotbind turn an OpenAPI specification into a hosted MCP endpoint?

Yes. In the Business Build dashboard, upload OpenAPI 3.0/3.1 JSON, preview selected tools, download schemas or an executable server package, and explicitly deploy a private hosted endpoint. Owners can confirm a Sandbox and enroll validated read-only Checks. Credentials are encrypted; client-credentials OAuth and temporary operation grants are supported. Writes require opt-in. Local verification is not real-assistant or production certification.

How does the Gateway help agents understand errors?

Pro and Business Owners can enable structured error hints per Gateway key. Hints use HTTP status and fields explicitly identified by validation errors, with Retry-After guidance where provided. They never invent renamed fields or refresh URLs, expose raw credential-bearing errors, or automatically retry writes. Successful responses and streams are unchanged.

Can we delegate API operations without sharing our upstream credentials?

Business Owners can store encrypted, operation-bound credentials and consent to exact tools for an API key or OAuth client/member for 15-60 minutes. Vended tokens authorize guarded execution, not raw-secret access. Credential rotation, expiry and revocation invalidate grants; interactive upstream login and 2FA remain human-owned.

Can our data stay in the EU?

Yes. You pick the hosting region: EU hosting runs the gateway, the monitoring and all stored data on infrastructure in the European Union with a GDPR data processing agreement (included in Business), or US hosting on infrastructure in the United States under standard contract terms.

Are caching, fallback and guardrails included in the Gateway?

No. Pro and Business include the Gateway itself: one OpenAI-compatible endpoint, an access key per app, a monthly budget, a request rate and a per-key usage log. Semantic Cache ($19), Content Guardrails ($19), Request Analytics ($15), Model Failover ($29) and Advanced Rate Limits ($12) are optional monthly modules. Turning a module on does not create a fallback rule or an extra quota until you add one.

// contact

Tell us what your customers should be able to ask.

Send a message. We reply with what to connect, which assistants to target, and a fixed price. No call.