MCP Server & Agent API

Most content APIs hand you a model and leave the hard parts — strategy, footage, captions, music licensing, scheduling — to you. This one hands you the finished artefact. Behind the endpoint is the same pipeline that runs the platform's own customers: a researcher that builds a content calendar from the brand's niche, a producer that renders 9:16 video, and a caption writer that emits a variant for all 11 platforms whether or not the brand has connected them.

The interface is deliberately small. Seventeen tools, one webhook event, one frozen package schema. Nothing expensive happens before payment: creating a brand writes a row and stops, so an agent that provisions a hundred brands and abandons them costs a hundred rows.

Today the server is deliver-only: we produce and hand over, and your agent publishes wherever it likes. Publish-through — approving a post and having us push it to the brand's connected accounts — is not yet exposed.

Quickstart

Every call below works against the live server. Replace the placeholders and run them in order; the whole funnel is nine calls.

1. 1 Register

   Unauthenticated. The key is returned in the response body and never sent by email — no email we send ever contains a credential.

   curl -X POST https://api.myinfluenceai.com/mcp/register \
     -H 'Content-Type: application/json' \
     -d '{
       "name": "Acme Agent",
       "contact_email": "[email protected]",
       "accepted_terms": true
     }'

2. 2 Verify the contact address

   The key is live from the moment it is issued, but crippled: until you click the link we email, it reaches exactly three tools. Lost the mail? POST /mcp/register/resend with {"agent": "<slug>"}.

3. 3 Connect an MCP client

   Streamable HTTP, stateless. Point any MCP-capable client — Claude, an SDK client, the MCP Inspector — at the endpoint with your key as a bearer token.

   {
     "mcpServers": {
       "myinfluenceai": {
         "type": "http",
         "url": "https://api.myinfluenceai.com/mcp",
         "headers": {
           "Authorization": "Bearer mi_live_<key_id>_<secret>"
         }
       }
     }
   }

4. 4 Find out what you hold

   whoami answers "what am I allowed to do" in one call — principal, scopes, rate limit, verification state and the brands this key can reach. Call it first whenever something is denied and you do not know why. platform_capabilities returns the formats, pillars and caption platforms, so nothing has to be hardcoded.

5. 5 Analyse a website and pick a plan

   website_analyze reads a public URL and proposes brand settings — niche, audience, tone, keywords. plan_list returns the plans and how much they deliver per day. Both are read-only.

6. 6 Create the brand

   brand_create writes the brand row, grants it the MCP add-on and records the reseller consent — then stops. No research, no production, no spend. You must pass accepted_terms: true having actually read the linked documents on the brand's behalf.

7. 7 Send a human to checkout

   checkout_create returns a Stripe URL. An agent cannot pay. A person completes that checkout, Stripe calls us, and the brand activates. Until it does, every brand-scoped tool refuses with entitlement_required.

8. 8 Configure delivery and prove your verifier works

   delivery_configure takes your webhook URL and returns a signing secret once. Store it before you read the next line. Then delivery_test sends a signed sample so you can confirm your HMAC check accepts a good signature — and, more importantly, rejects a bad one — before any real content depends on it.

9. 9 Receive the daily package

   From then on, each produced item is POSTed to your endpoint as a package.ready event. If your endpoint was down, package_list and package_get return the same bytes, and delivery_replay re-sends.

Authentication

Keys look like mi_live_<key_id>_<secret> and go in an Authorization: Bearer header. The secret half is shown once, at creation, and is stored only as a peppered hash — we cannot recover it for you, so a lost key is replaced, not retrieved.

Self-serve registration always grants the same thing: the developer tier and the provision preset. It ignores any tier, scope or brand grant you ask for. A self-serve key can never carry publish:write, outreach:send or pii:read — those take an irreversible or privacy-bearing action on someone's behalf, so granting them is a separate, deliberate act.

tools/list is filtered to what your key can actually call. A tool you would only ever be denied is not advertised, because it would waste your context and your calls.

Before your email is verified

Registration is unauthenticated, so a contact address is an unproven claim until someone opens the link we send. An unverified key reaches only the tools listed below; everything else returns email_verification_required. The key is issued before verification and crippled, rather than minted on click, so a live credential never has to travel through an inbox.

Scopes

A key carries an explicit, literal list of scopes. Presets are a shortcut at minting time only — the stored key always holds the expanded list, so widening a preset later can never widen a key that already exists.

Presets

• readonly — brands:read, content:read, analytics:read, social:read
• deliver — brands:read, content:read, delivery:write
• provision — brands:read, content:read, billing:write, delivery:write

Self-serve registration always grants provision, and the scopes marked above are never included in any preset and never implied by another scope.

Tool reference

All 17 tools, generated from the running registry. Brand-scoped tools take a brand slug and require both the MCP add-on and an active subscription on that brand; UUIDs are refused. Writes change state and are audited.

{
  "type": "object",
  "properties": {
    "brand": {
      "type": "string",
      "description": "Brand slug."
    }
  },
  "required": [
    "brand"
  ],
  "additionalProperties": false
}

{
  "type": "object",
  "properties": {
    "brand_name": {
      "type": "string",
      "maxLength": 120,
      "description": "Display name. The slug is derived from it unless you pass one."
    },
    "slug": {
      "type": "string",
      "maxLength": 63,
      "pattern": "^[a-z0-9][a-z0-9-]*[a-z0-9]$",
      "description": "Optional explicit slug."
    },
    "website": {
      "type": "string",
      "maxLength": 2000
    },
    "niche": {
      "type": "string",
      "maxLength": 500
    },
    "tone": {
      "type": "string",
      "maxLength": 200
    },
    "plan": {
      "type": "string",
      "maxLength": 40,
      "description": "Intended plan id; used by checkout_create."
    },
    "owner_email": {
      "type": "string",
      "maxLength": 320,
      "description": "Optional. The end brand’s email, for the claim invite and consent record."
    },
    "accepted_terms": {
      "type": "boolean",
      "description": "Must be true. You are asserting authority to accept on the brand’s behalf."
    }
  },
  "required": [
    "brand_name",
    "accepted_terms"
  ],
  "additionalProperties": false
}

{
  "type": "object",
  "properties": {
    "brand": {
      "type": "string",
      "description": "Brand slug."
    }
  },
  "required": [
    "brand"
  ],
  "additionalProperties": false
}

{
  "type": "object",
  "properties": {},
  "additionalProperties": false
}

{
  "type": "object",
  "properties": {
    "brand": {
      "type": "string",
      "description": "Brand slug."
    },
    "plan": {
      "type": "string",
      "maxLength": 40,
      "description": "Plan id from plan_list. Defaults to the brand’s current plan."
    },
    "success_url": {
      "type": "string",
      "maxLength": 2000
    },
    "cancel_url": {
      "type": "string",
      "maxLength": 2000
    }
  },
  "required": [
    "brand"
  ],
  "additionalProperties": false
}

{
  "type": "object",
  "properties": {
    "brand": {
      "type": "string",
      "description": "Brand slug."
    },
    "webhook_url": {
      "type": "string",
      "maxLength": 2000,
      "description": "Public https:// endpoint. Private and loopback addresses are refused."
    },
    "mode": {
      "type": "string",
      "enum": [
        "deliver_only",
        "publish_through"
      ],
      "description": "deliver_only is the Phase 1 product. publish_through is not yet available."
    },
    "platforms": {
      "type": "array",
      "description": "Caption variants to include. Omit for all 11."
    },
    "telegram_enabled": {
      "type": "boolean",
      "description": "Also send the Telegram approval card. Default false."
    },
    "paused": {
      "type": "boolean",
      "description": "Stop delivering without losing configuration."
    },
    "rotate_secret": {
      "type": "boolean",
      "description": "Mint a new signing secret. Breaks existing verification until you redeploy."
    }
  },
  "required": [
    "brand"
  ],
  "additionalProperties": false
}

{
  "type": "object",
  "properties": {
    "brand": {
      "type": "string",
      "description": "Brand slug."
    },
    "delivery_id": {
      "type": "string",
      "maxLength": 40,
      "description": "From package_list."
    }
  },
  "required": [
    "brand",
    "delivery_id"
  ],
  "additionalProperties": false
}

{
  "type": "object",
  "properties": {
    "brand": {
      "type": "string",
      "description": "Brand slug."
    }
  },
  "required": [
    "brand"
  ],
  "additionalProperties": false
}

{
  "type": "object",
  "properties": {
    "brand": {
      "type": "string",
      "description": "Brand slug."
    },
    "job_id": {
      "type": "integer",
      "description": "Job id returned by a mutating tool."
    }
  },
  "required": [
    "brand",
    "job_id"
  ],
  "additionalProperties": false
}

{
  "type": "object",
  "properties": {
    "brand": {
      "type": "string",
      "description": "Brand slug."
    },
    "request_id": {
      "type": "string",
      "maxLength": 120,
      "description": "From a package.ready payload or pipeline_status."
    }
  },
  "required": [
    "brand",
    "request_id"
  ],
  "additionalProperties": false
}

{
  "type": "object",
  "properties": {
    "brand": {
      "type": "string",
      "description": "Brand slug."
    },
    "status": {
      "type": "string",
      "enum": [
        "pending",
        "sending",
        "delivered",
        "failed",
        "dead"
      ]
    },
    "since": {
      "type": "string",
      "maxLength": 40,
      "description": "ISO 8601 timestamp."
    },
    "limit": {
      "type": "integer",
      "minimum": 1,
      "maximum": 100,
      "description": "Default 20."
    },
    "cursor": {
      "type": "string",
      "maxLength": 40,
      "description": "next_cursor from a previous page."
    }
  },
  "required": [
    "brand"
  ],
  "additionalProperties": false
}

{
  "type": "object",
  "properties": {
    "brand": {
      "type": "string",
      "description": "Brand slug."
    }
  },
  "required": [
    "brand"
  ],
  "additionalProperties": false
}

{
  "type": "object",
  "properties": {},
  "additionalProperties": false
}

{
  "type": "object",
  "properties": {},
  "additionalProperties": false
}

{
  "type": "object",
  "properties": {
    "brand": {
      "type": "string",
      "description": "Brand slug."
    }
  },
  "required": [
    "brand"
  ],
  "additionalProperties": false
}

{
  "type": "object",
  "properties": {
    "website": {
      "type": "string",
      "maxLength": 2000,
      "description": "Public https:// URL of the brand’s site."
    }
  },
  "required": [
    "website"
  ],
  "additionalProperties": false
}

{
  "type": "object",
  "properties": {},
  "additionalProperties": false
}

Errors

A tool failure comes back as a normal result marked isError, not a protocol fault, so your agent can read it and correct itself. Every error carries a code, a message and a next_step written for a program rather than a person.

brand_not_found and brand_not_authorized return the same message on purpose: a key must not be able to discover which brands exist by probing.

Webhook delivery

Every delivery carries an X-MI-Signature header in the form t=<unix>,v1=<hex>, where v1 is an HMAC-SHA256 of the literal string "<t>.<raw body>" under your signing secret. The timestamp is inside the signed value, so a captured body cannot be replayed with a fresh timestamp.

Three things a correct verifier does, and a plausible one does not: hash the raw request body (not a re-serialised object — key order changes the bytes), compare with a constant-time comparison, and reject anything outside a ±5 minute window. A verifier that skips the timestamp check silently accepts forged calls forever.

We accept any 2xx as delivered. Anything else retries on a fixed schedule and then stops. Respond fast and process asynchronously: a slow endpoint is indistinguishable from a broken one.

Headers

POST /your/webhook HTTP/1.1
Content-Type: application/json
X-MI-Event: package.ready
X-MI-Signature: t=1756512000,v1=9f86d081884c7d659a2feaa0c55ad015a3bf4f1b2b0b822cd15d6c15b0f00a08

Verify it — Node

import crypto from 'crypto';

// body must be the RAW request bytes, not a re-serialised object:
// JSON.stringify(JSON.parse(body)) can reorder keys and change the hash.
export function verify(rawBody, header, secret, toleranceSeconds = 300) {
  const parts = Object.fromEntries(
    String(header || '').split(',').map(kv => kv.split('=').map(s => s.trim())),
  );
  const t = parseInt(parts.t, 10);
  if (!Number.isFinite(t)) return false;

  // Without this, a captured body replays forever.
  if (Math.abs(Math.floor(Date.now() / 1000) - t) > toleranceSeconds) return false;

  const expected = crypto.createHmac('sha256', secret)
    .update(`${t}.${rawBody}`, 'utf8').digest('hex');
  const a = Buffer.from(expected, 'hex');
  const b = Buffer.from(String(parts.v1 || ''), 'hex');
  if (a.length !== b.length || a.length === 0) return false;
  return crypto.timingSafeEqual(a, b);
}

Verify it — Python

import hashlib, hmac, time

def verify(raw_body: bytes, header: str, secret: str, tolerance_seconds: int = 300) -> bool:
    parts = dict(
        kv.strip().split("=", 1) for kv in (header or "").split(",") if "=" in kv
    )
    try:
        t = int(parts["t"])
    except (KeyError, ValueError):
        return False

    # Without this, a captured body replays forever.
    if abs(int(time.time()) - t) > tolerance_seconds:
        return False

    signed = f"{t}.".encode() + raw_body
    expected = hmac.new(secret.encode(), signed, hashlib.sha256).hexdigest()
    return hmac.compare_digest(expected, parts.get("v1", ""))

Retries

A non-2xx response (or a timeout) is retried 6 times in total, then marked dead:

1m → 5m → 30m → 2h → 6h → 24h → dead

That is roughly 33 hours of tolerance for an endpoint that is down. A dead delivery is not lost — package_list still returns it and delivery_replay re-sends it.

The daily package

One frozen shape, assembled once and used identically by the webhook push and the package_get pull — so a consumer catching up after an outage gets exactly what it normally receives.

The one thing consumers get wrong: media.kind is "video" or "carousel". For a carousel, video_url is null and the ordered slide URLs are in media.slides. Branch on kind; a downloader that assumes a single URL breaks on the first carousel day.

Music credit is not optional. When music.credit_required is true, the ready-made music.credit_line must appear in the caption you publish — not in file metadata. Instagram, TikTok and YouTube all re-encode uploads and strip container tags, so metadata attribution is not attribution. Nearly every track in the library is CC BY.

Media URLs expire. expires_at is on every package: fetch and store the bytes, do not hotlink.

Captions

A variant is generated for all 11 platforms on every post, regardless of what the brand has connected, so a deliver-only consumer can publish anywhere: linkedin, tiktok, x, reddit, instagram, facebook, youtube, bluesky, pinterest, threads, mastodon. Narrow the set with the platforms argument to delivery_configure. A platform with no stored caption is absent rather than null, so "not generated" is distinguishable from "generated empty".

Shape

A real payload, assembled by the same code that builds yours. This one was configured for three platforms so the sample stays readable; by default captions holds all 11.

{
  "package_version": "1",
  "brand": "acme-dental",
  "request_id": "req_9f4c21",
  "date": "2026-08-30",
  "format": "talking_points",
  "pillar": "educational",
  "hook": "Three things every new patient asks — answered in 30 seconds.",
  "media": {
    "kind": "video",
    "video_url": "https://cdn.my-influence-ai.com/v/req_9f4c21.mp4",
    "slides": null,
    "thumbnail_url": "https://cdn.my-influence-ai.com/v/req_9f4c21.jpg",
    "duration_seconds": 28
  },
  "captions": {
    "linkedin": {
      "text": "Patients ask the same three questions before every first appointment. Here is how we answer them.",
      "hashtags": []
    },
    "tiktok": {
      "text": "The 3 questions we get every single day",
      "hashtags": [
        "#dentaltok",
        "#patienttips"
      ]
    },
    "instagram": {
      "text": "Three things every new patient asks. Save this for later 🦷",
      "hashtags": [
        "#dentalcare",
        "#patienttips"
      ]
    }
  },
  "default_caption": "Three things every new patient asks. Save this for later.",
  "music": {
    "title": "Sunlit",
    "artist": "Kirkoid",
    "license": "CC BY 3.0",
    "license_url": "https://creativecommons.org/licenses/by/3.0/",
    "credit_required": true,
    "credit_line": "🎵 Music: \"Sunlit\" by Kirkoid (ccMixter, CC BY 3.0)"
  },
  "expires_at": "2026-09-29T06:00:00.000Z"
}

Rate limits and quotas

Rate limits are per key and reported by whoami; a rate_limited error carries retry_after_seconds. Back off exponentially — a retry loop against a limit is the fastest way to get a key disabled.

Plan quotas are enforced per brand and fail closed: if we cannot determine your remaining quota, the call is refused rather than allowed. That is deliberate for an automated caller, and it is why quota_unavailable is a distinct code from quota_exceeded — the first is transient and worth retrying, the second is not.

Every response carries a request_id. Quote it when you contact support; it is how we find your call.

Versioning

The tool schemas and the package schema are frozen contracts, checked against a fixture in CI — an accidental change to either fails our build, not yours.

Additive fields do not bump a version. A breaking change mounts a new path (/mcp/v2); /mcp never silently becomes v2. Deprecations run 90 days, are flagged on the affected responses, and are emailed to the owners of affected keys.

Machine-readable discovery lives at GET https://api.myinfluenceai.com/mcp/info and https://agents.myinfluenceai.com/.well-known/mcp/server.json; both are unauthenticated and always describe the deployed server.

Terms

Creating a brand means accepting these on that brand's behalf. Read them before you pass accepted_terms: true.

• Terms of Service
• Privacy Policy
• Acceptable Use Policy
• Data Processing Addendum
• Subprocessors

Frequently asked questions