MCP server - SideShift

SideShift exposes a remote Model Context Protocol server, so an AI agent can operate a SideShift company directly - listing campaigns, reading posts, creating contracts, running payouts - instead of you writing HTTP calls. There is nothing to download. The server is hosted at https://app.sideshift.app/api/mcp and you connect a client to it.

What you get

The MCP server fronts the same OAuth-scoped capabilities as the REST API. Every capability scope maps to tools, so the tool set mirrors the REST surface, plus a whoami tool for checking which company, scopes, and sub-account access a token is bound to.

Read scopes expose read-only tools
Write and sensitive tools carry a confirmation guardrail in their description
The claim-link capability is available as create_quickpay_claim_link

Before you start

You need three things.

An MCP-capable client

Claude Desktop, Claude Code, Cursor, VS Code, or any client that speaks MCP over Streamable HTTP.

A registered OAuth client

The MCP server uses the same OAuth 2.1 flow as the REST API - there is no MCP-specific auth scheme. Register via Dynamic Client Registration to get a client_id. See Platform API.

An access token

Run the authorization-code flow with PKCE (S256) to get a bearer token. Tokens minted for MCP are multi-audience, so one token works for both /api/oauth/v1 and /api/mcp - you do not need a separate credential.

If a request arrives unauthorized, the server replies 401 with a WWW-Authenticate challenge pointing at the RFC 9728 protected-resource metadata. Compliant clients use this to discover the authorization server automatically, so many clients can complete setup from the URL alone.

Connect a client

Claude Code
Claude Desktop
Cursor
VS Code

claude mcp add --transport http sideshift https://app.sideshift.app/api/mcp

Claude Code will walk you through the OAuth flow on first use.

Settings → Connectors → Add custom connector, then enter:

https://app.sideshift.app/api/mcp

Approve the SideShift consent screen when it appears, choosing the company and scopes to grant.

Add to ~/.cursor/mcp.json:

{
  "mcpServers": {
    "sideshift": {
      "url": "https://app.sideshift.app/api/mcp"
    }
  }
}

Add to .vscode/mcp.json in your workspace:

{
  "servers": {
    "sideshift": {
      "type": "http",
      "url": "https://app.sideshift.app/api/mcp"
    }
  }
}

Two different screens are called Connectors. Claude Desktop’s Settings → Connectors points an outside AI client at SideShift - that is this page. Inside the SideShift app, Settings → Sidekick Connectors goes the other direction: Sidekick acts as the MCP client and connects out to external vendors’ MCP servers. See Sidekick connectors.

Setup snippets follow each client’s documented format for remote HTTP MCP servers. Client configuration changes often - if one is rejected, check that client’s current docs. The server URL and the OAuth flow are the parts that come from the SideShift specification.

Team members, scoped Platform API keys, OAuth grants, and MCP sessions share the same public capability names. A saved permissions.scopes array is the exact grant - an empty array grants nothing; a membership without a scope array keeps its prior behavior, including the legacy reporting-only Employee ceiling, until an explicit selection is saved. Every verification re-intersects the session’s scopes with the subject’s current membership, so a refresh can narrow access but never exceed the originally consented grant. See Team role ceiling for human grants.

Using the tools

Once connected, confirm which company you’re operating on before doing anything else:

Use the whoami tool to show which SideShift company and scopes I'm connected with.

Then work in plain language:

List my active campaigns and show how many creators applied to each.

Every request is bound to the tenant and the token’s effective scopes, so an agent can only do what it was authorized to do. Requesting a capability outside those scopes fails rather than silently doing something else. For a human grant, that effective scope set can be narrower than what was requested at consent. The ceiling is checked live, while the grant retains the consented set: a later authorized role or permission change can take effect on refresh, but cannot add access that was never consented to. See the note above. Changing the dashboard company does not retarget an existing MCP connection. To switch the connected company, disconnect and authorize again, selecting the intended company. Disconnecting revokes the old grant; reconnecting does not revive old tokens. For the manager and teammate flows, see Team and integration permissions.

Agency sub-account delegation

An agency’s OAuth grant can extend to its sub-accounts. At consent, the user grants access to all sub-accounts (including ones created later) or an explicit current subset. Call whoami to see what was granted: its subaccountAccess field is null when no sub-account access exists, "all" for every sub-account, or an explicit list of sub-account company ids. Every consolidated business tool accepts an optional act_as_subaccount_id argument (help_query is the exception — it stays tenant-free). Pass a sub-account’s company id to run that call against the sub-account instead of the connected parent; resolve ids with list_companies first rather than guessing. Omit the argument to act on the connected account itself. The executor maps it to an X-Act-As-Company header on the underlying request. Each delegated call re-verifies the current parent relationship, the user’s own membership, and the target sub-account’s scope ceiling - the consented subaccountAccess set is a ceiling, not a standing grant, so a child removed from the agency or a permission change on either side takes effect on subsequent verified requests. A narrower direct child membership takes precedence over inherited parent access. Aggregate reads across sub-accounts filter to that same live-permitted set. See Agency access for setup and error recovery. Platform API keys support the same REST delegation; remote MCP connections use OAuth.

Finding the right tool

Some hosts cap how many tools actually reach the model. SideShift registers six discovery tools first and core product tools immediately after, so the important actions survive both insertion-order and alphabetical truncation while uncapped clients still get the full typed surface. When an action is not obviously available, discover it rather than concluding it does not exist:

Search by intent

Call find_capability with the user’s ordinary-language request. The older catalog_search name performs the same read.

Load the exact schema

Call catalog_get_schema for the selected capability, especially before a configurable write.

Gather and confirm

Collect every required value, present enum choices, and confirm the payload with the user.

Invoke at the matching risk level

Call catalog_invoke_read, catalog_invoke_write, or catalog_invoke_sensitive.

Search results report whether the current grant holds the required scope, so a missing scope leads to reauthorization rather than a false “unsupported” answer. whoami reports both the direct and complete-catalog tool counts.

Product-language names such as create_job, update_job, and save_discover_offer are canonical. The older upsert_job and upsert_discover_offer names stay callable for existing integrations but are suppressed from search results.

Working with marketplace jobs

Task	Call	Notes
Create or post a job	`create_job`	Gather title, description, category, cover image, and the complete payment configuration first. A mixed payment needs an explicit fixed-vs-per-post base. Spends one posting credit, so confirm before calling.
Edit an existing job	`update_job`	Resolve the exact owned job with `list_jobs` or `get_job` first, then patch only the requested fields. Does not spend a credit. Changing the payment model requires its complete matching configuration in the same patch.
Archive, unarchive, delete	`update_job_status`	Exclusive to this tool, which enforces capacity, refund, feed, and audit invariants.
Post a campaign to the marketplace	`create_job` with the campaign’s `programId`	Create or resolve the campaign first. `create_campaign` has no `marketplace_job` field.

Named skills

The MCP server exposes named skills: curated multi-step procedures. Each skill is available two ways:

As an MCP prompt - one per built-in skill, named exactly as in the table below - so hosts that support prompts can show them in their slash menus. Selecting a prompt injects the skill body as a user message.
Via the invoke_skill tool (campaigns:read, read). Pass the skill name (for example post-job). The tool description lists every skill and its trigger rules; unknown names return the valid name list. Bundled instructions only - no account data.

Skill name	Title	Use when
`boost-post`	Boost a post	Turn a winning organic post into a paid ad on TikTok (Spark) or Meta (creator partnership/whitelist, or the brand’s own Page); boost, promote, amplify, or whitelist a post. Launching can start real ad spend and code requests message real creators, so both need explicit confirmation. Not for organic analytics or for managing an existing ad’s budget.
`create-campaign`	Create a campaign	Design/launch a campaign from proven templates; campaign setup or “just set it up for me”. Not for editing an existing campaign or posting a marketplace job alone.
`find-creators`	Find creators	Search the creator marketplace, shortlist candidates, and invite them to a job; “who should I work with”, candidates matching a niche, rate, or platform. Not for people who already applied, for offering a contract to a known creator, or for roster questions.
`fund-wallet`	Fund the wallet	Wallet balance, adding or topping up funds, “do we have enough to pay creators”, an insufficient-balance error, or moving money between agency and client wallets. Adding new funds is browser-only via a handoff link. Not for actually paying creators or for what SideShift charges the company.
`grill-my-brief`	Grill my brief	User pastes a UGC/creator/ad brief and asks to grill, review, audit, critique, or stress-test it. Not for writing briefs from scratch, creator outreach copy, paid-ad copy, or general copywriting feedback.
`hire-creator`	Hire a creator	Hire, sign, or contract a named creator, make them an offer, add a creator to a campaign, or send custom per-creator terms. Not for finding candidates, for approving a campaign application (approval IS the contract), or for shareable join links on their own.
`invoice-client`	Invoice a client	Invoice or bill the company’s own customer, check whether an invoice was paid, chase or void it, settle a completed checkout, send receipts, or refund. Not for what SideShift charges the company, for paying people, or for creator payouts.
`message-creators`	Message creators	Message, DM, broadcast, announce, remind, or nudge creators - one person, a campaign roster, or a saved group. Every send reaches a real person and needs the exact approved text and recipient list. Not for recruiting outreach to non-roster creators or for messaging job applicants about an application.
`performance-report`	Performance report	How things are going, a weekly or monthly report, recap, or summary; campaign health, pacing, benchmarks, top performers, or what needs fixing. Not for one quick number or chart, and not for moderating content.
`post-job`	Post a job	Post/publish a Creator Marketplace job (spends 1 job credit; confirm complete posting first). Not for editing an existing job, creating the campaign itself, or reviewing applicants.
`quickpay`	Quick Pay someone	Pay someone outside the contract system - a one-off payment by email, a staged batch, a recurring schedule, or a claim link; “send $X to …”. Every rail moves real wallet money and claim links debit at creation, so recipient and amount need explicit confirmation. Not for contract payouts or for collecting money via a customer invoice.
`resolve-dispute`	Resolve a dispute	A creator-payment dispute: review the claim and evidence, then pay, pay-and-remove, or contest with counter-evidence; also “what disputes are open”. Paying settles with real money, and the creator is notified whichever way it goes. Not for ordinary owed payouts or for refunding a customer invoice.
`review-applicants`	Review applicants	Review, compare, accept, or decline applicants and applications - marketplace job applicants, campaign handle requests, and inbound creator invites; “who applied”, or a pending backlog. Approving a campaign application CREATES the contract, so every approval needs explicit confirmation. Not for finding new creators or for messaging the roster.
`review-content`	Review content	Review, approve, or reject pending video submissions and tracked posts; “what needs review”, requesting revisions, or excluding a bad post from analytics or payouts. Rejections message the creator, so each needs a confirmed reason. Not for performance analysis or for paying for approved content.
`run-payouts`	Run payouts	Pay creators, run or send payouts, “what do we owe”, review pending payouts, skip a cycle, or fix an amount before paying. Executing moves real, irreversible money and always needs a dry-run preview plus explicit approval. Not for one-off payments to a single person or for payment disputes.
`write-brief`	Write a brief	Write, draft, or clarify a UGC creator brief or marketplace job description, grounded in the campaign’s saved fields, the brand’s profile, and top-performing briefs; what should go in a creator brief. Not for auditing a brief that already exists or for campaign setup.

Tool catalog

A curated highlight of the tools an agent reaches for most often. The server exposes far more than this: see the capability reference for all of them, generated from the live registry. Risk levels: read makes no writes, write mutates SideShift state reversibly, and sensitive has real money, email, or direct-message side effects, so it is sandbox-rejected and blocked under MCP_SAFE_MODE.

Tool	Scope	Risk	What it does
`whoami`	-	read	Company id, granted scopes, client id, sandbox flag, sub-account access
`find_capability`	-	read	Find the right capability from an ordinary-language request
`catalog_search`	-	read	Search the complete catalog by user intent
`catalog_get_schema`	-	read	Load one exact authoritative input schema
`catalog_invoke_read`	capability’s scope	read	Invoke a discovered read safely
`catalog_invoke_write`	capability’s scope	write	Invoke a confirmed discovered write
`catalog_invoke_sensitive`	capability’s scope	sensitive	Invoke a confirmed sensitive action with all existing backstops
`invoke_skill`	campaigns:read	read	Load a named SideShift skill: a multi-step procedure to follow for the rest of the turn.
`list_campaigns`	campaigns:read	read	List campaigns (filter by status/search)
`get_campaign`	campaigns:read	read	Fetch one campaign
`create_campaign`	campaigns:write	write	Create a campaign
`update_campaign`	campaigns:write	write	Update name/emoji/description
`archive_campaign`	campaigns:write	write	Archive a campaign
`duplicate_campaign`	campaigns:write	write	Clone a campaign into a new one
`add_campaign_ghost_handle`	campaigns:write	write	Add one tracked handle to a campaign
`bulk_add_campaign_ghost_handles`	campaigns:write	write	Add up to 500 tracked handles to a campaign
`add_campaign_ghost_video`	campaigns:write	write	Add one tracked video to a campaign
`bulk_add_campaign_ghost_videos`	campaigns:write	write	Add up to 500 tracked videos to a campaign
`import_campaign_analytics_history`	campaigns:write	write	Validate or import observed pre-tracking daily history
`get_payment_structure`	campaigns:read	read	Read a campaign’s payout model
`create_job`	jobs:write	write	Create/post a Creator Marketplace job (spends one posting credit)
`update_job`	jobs:write	write	Partially update an owned job without spending a posting credit
`list_jobs`	jobs:read	read	List the connected company’s jobs
`get_job`	jobs:read	read	Fetch one owned job
`list_applications`	applications:read	read	List creator applications
`get_application`	applications:read	read	Fetch one application
`update_application_status`	applications:write	write	Approve / reject an application
`list_creators`	creators:read	read	List creators (filter by campaign/search)
`get_creator`	creators:read	read	Fetch one creator
`list_creator_collections`	creators:read	read	List saved creator collections
`create_creator_collection`	creators:write	write	Create a creator collection
`create_campaign_invite`	creators:write	write	Create a shareable campaign invite link
`list_contracts`	contracts:read	read	List contracts (filter by status/campaign/creator)
`get_contract`	contracts:read	read	Fetch one contract
`create_contract`	contracts:write	write	Offer a contract to a creator (notifies them)
`update_contract`	contracts:write	sensitive	Change one creator’s base retainer and/or posting platforms; regenerates the PDF and notifies on re-review
`cancel_contract`	contracts:write	write	Cancel a contract
`save_contract_signature`	contracts:sign	sensitive	Upload and save an e-signature (drawn, typed, or image) for a contract
`list_posts`	posts:read	read	List tracked posts (filter by campaign/platform/creator/date)
`get_post`	posts:read	read	Fetch one post + latest metrics
`get_post_metrics`	posts:read	read	A post’s daily metrics time-series
`export_posts`	posts:read	read	Export posts as CSV (filtered / by ids)
`get_analytics_accounts`	analytics:read	read	Read tracked-account analytics
`list_payouts`	payouts:read	read	List historical payouts + summary
`list_pending_payouts`	payouts:read	read	List pending/owed payouts + summary
`get_payout_stats`	payouts:read	read	Wallet balance + lifetime payout summary
`execute_payouts`	payouts:write	sensitive	Execute (or dry-run) contract payouts, moves real money
`quick_pay`	payouts:write	sensitive	One-off quick payment to a creator, moves real money
`list_invoices`	invoices:read	read	List invoices (filter by status/customer)
`get_invoice`	invoices:read	read	Fetch one invoice
`create_invoice`	invoices:write	sensitive	Create an invoice (may email the customer)
`send_invoice`	invoices:write	sensitive	Email an existing invoice
`void_invoice`	invoices:write	sensitive	Void an existing invoice
`list_inbox`	messages:read	read	The company’s merged inbox: start here for “show my conversations”
`list_conversations`	messages:read	read	List the company’s DM conversations only (not the inbox)
`get_messages`	messages:read	read	List messages in a DM conversation
`list_support_channels`	messages:read	read	List support channels, the main employer-creator conversation type
`get_support_channel_messages`	messages:read	read	List messages in a support channel
`create_conversation`	messages:write	sensitive	Create a 1:1 or group conversation
`send_message`	messages:write	sensitive	Send a direct message to a creator
`send_support_channel_message`	messages:write	sensitive	Send a message in an existing support channel
`react_to_message`	messages:write	sensitive	Add an emoji reaction to a message
`forward_message`	messages:write	sensitive	Forward a message’s content to the conversation’s other member
`add_conversation_member`	messages:write	sensitive	Add a member to a group conversation
`leave_conversation`	messages:write	sensitive	Leave a group conversation as the company
`rename_conversation`	messages:write	sensitive	Rename a group conversation
`message_applicant`	applicants:write	sensitive	Message a creator who applied to a job
`resolve_applicant_support_channel`	applicants:write	sensitive	Create or reuse the support channel for an applicant
`send_discover_message`	discover:write	sensitive	Send a message to a Discover agency
`get_settings`	settings:read	read	Read API-visible company settings
`get_company`	settings:read	read	Read the company profile
`update_settings`	settings:write	sensitive	Update allowlisted company settings
`list_invites`	creators:read	read	List campaign invite links
`revoke_invite`	creators:write	write	Revoke a campaign invite link

create_invite is deliberately not exposed; use create_campaign_invite, which is the same capability.

Confirm intent before any sensitive call, and prefer execute_payouts with dry_run: true before moving money. update_contract counts as sensitive because a real edit sends the creator a review link, though an already-current no-op sends nothing.

Companies on SideShift messaging

SideShift moved chat from a connected provider to its own messaging, one company at a time. A company moves the first time a person opens the app. For a company that has moved, the messaging tools above are served from native rooms:

Every conversation, channel and message id is a UUID. An older conversation id or feed_… support-channel id still resolves to its room.
Rows carry chatType: "native", messages carry source: "native", and every write answers lane: "native".
attachment_ids are provider file ids and are refused on a native room (400 invalid_request, reason attachments_unsupported_on_native_conversation).
While the move runs, every messaging tool answers 409 conflict with meta.reason: "messaging_migrating". Nothing is sent. Retry in a minute.

The messaging recipe shows the response shapes.

Safety rules

These are enforced server-side, not by the client.

Rule	Behavior
Sandbox grants	Sensitive tools - anything that moves money or triggers outbound side-effects - are rejected for test and sandbox grants
Global kill switch	`MCP_SAFE_MODE` disables sensitive tools server-wide
Idempotency	Money-moving operations require an `Idempotency-Key`, so a retried call never double-charges
Scope binding	One effective company per call; consented agency parents may select one direct child with `act_as_subaccount_id`

payouts:write and settings:write are flagged sensitive in the specification because they move money and change company configuration. Grant them only when an integration genuinely needs them.

Tool errors

Tool errors are actionable and map from the oauth/v1 error envelope. The full code and status registry is on the Errors, rate limits, and idempotency page.

Condition	What you get back
Missing scope	”requires the ‘X:write’ scope …” in the message. The attached OAuth challenge names the full scope set to request on reconnect: every scope you already hold, plus ‘X:write’. It does not name only ‘X:write’.
Unknown or cross-tenant id	”Not found … or it is not visible to this company.”
Sensitive tool on a sandbox token	”performs a real side effect and is blocked for sandbox tokens.”
Sensitive tool under `MCP_SAFE_MODE`	”disabled because the server runs in MCP_SAFE_MODE.”
Rate limited	”Rate limit exceeded. Wait a few seconds and retry.”
Subscription inactive	”the company’s subscription is not active …”

Docs server

Separately from the product API above, this documentation site publishes its own MCP server so an agent can read these pages:

https://docs.sideshift.app/mcp

That one needs no authentication and grants no access to your SideShift data - it only reads documentation. You can also add it from the Copy page menu at the top of any page, which has one-click setup for Cursor and VS Code.

	Product server	Docs server
URL	`app.sideshift.app/api/mcp`	`docs.sideshift.app/mcp`
Auth	OAuth 2.1 required	None
Can do	Operate a company	Read these docs

​What you get

​Before you start

​Connect a client

​Using the tools

​Agency sub-account delegation

​Finding the right tool

​Working with marketplace jobs

​Named skills

​Tool catalog

​Companies on SideShift messaging

​Safety rules

​Tool errors

​Docs server

What you get

Before you start

Connect a client

Using the tools

Agency sub-account delegation

Finding the right tool

Working with marketplace jobs

Named skills

Tool catalog

Companies on SideShift messaging

Safety rules

Tool errors

Docs server