--- title: 'Meta Ads MCP: choose a server for your AI agent' description: 'Compare Meta Ads MCP servers: Meta’s official tools, Plainrouter approvals and independent counts, plus hosted and self-hosted options.' canonical: 'https://plainrouter.com/library/meta-ads-mcp-options' format: Comparison published_at: '2026-09-12' last_updated: '2026-10-29' --- # Meta Ads MCP: choose a server for your AI agent A Meta Ads MCP server connects an AI client to tools for reading and managing Meta advertising accounts through the Model Context Protocol. Choose Plainrouter when you want an approval queue for your agent’s Meta ad changes and your own arrival and conversion counts beside Meta’s figures; choose Meta’s official Ads MCP when its own reporting and ad-management tools cover your workflow. Published by Plainrouter, which is included in this comparison. All entries and linked pricing sources were checked on October 29, 2026. This is a documentation comparison, not a hands-on competitor benchmark; we have not tested competitors' authenticated products. The dated Plainrouter sandbox read below is historical evidence of synthetic connectivity only; that sandbox endpoint has since been retired. ## How we chose and checked these options We selected documented Meta Ads MCP options to compare the needs in the table below: platforms, hosting, change controls and independent conversion evidence. Provider documentation, public repositories and pricing pages were checked on October 29, 2026; an unverified capability is labelled accordingly rather than treated as absent. We have not tested competitors' authenticated products, and the dated Plainrouter sandbox read shown below is historical evidence of synthetic read connectivity, not an approval, production execution or a currently available test endpoint. Plainrouter publishes this comparison and is included under the same documentation-based checks for access, changes, hosting and price. ## Which Meta Ads MCP should you choose? For a Meta-focused team that wants to review changes and check Meta’s numbers against its own count, choose Plainrouter. If your main requirement is several ad platforms or a server you host yourself, choose around that requirement instead. | Your requirement | Choose | What you get and what to check | | --- | --- | --- | | Review Meta ad changes and compare your own count with Meta’s | Plainrouter | Ask holds every change for a person. Every change that runs gets a decision receipt. Plainrouter’s own arrivals and conversions sit beside Meta’s figures; this is not a measure of incrementality. | | Meta's own reporting and ad-management tools | Meta's official Ads MCP | Meta hosts reporting and ad-management tools. Its available server rules deny actions and constrain budget changes. | | Several advertising platforms | Pipeboard, Blend | Their published MCP products connect multiple ad platforms. Verify the selected platform's operations and account scope. | | A server you can operate and inspect yourself | GoMarble open source | Local Python server under the MIT license; the reviewed README lists reporting and object retrieval, not campaign-write tools. | | Other documented review workflows | AdKit, Markifact, NotFair | AdKit describes retained drafts; Markifact describes approval-based writes; NotFair describes staged, approval-gated mutations. Check the exact operation and where review happens. | For a client-neutral Plainrouter connection guide, use [hosted Meta Ads MCP setup](). For a wider category, see [AI tools for Facebook ads](). For Plainrouter in your terminal, follow the [Claude Code for Meta ads guide]() to connect with account access and approval controls. If you are choosing between Meta's own assistants, a client connected to its server and a separate approval workflow, compare [Meta AI ads tools](). ## What does Meta's official Ads MCP server do? Meta's server exposes reporting and ad-management tools over a hosted MCP endpoint. Its documented tool areas also include catalogs, signals and datasets, troubleshooting, A/B and lift tests, and activity logs. These are Meta's own business surfaces, not an independent revenue measurement system. [Meta overview](). Meta supports OAuth, user access tokens and Employee-role system-user access tokens; its setup page says Admin-role system-user tokens cannot access this server. The minimum token permissions are ` ads_mcp_management ` plus ` ads_read ` or ` ads_management `, with additional permissions enabling more tools. The same page distinguishes connecting with your own developer app from the AI-connector route that does not require owning one. Follow the route appropriate to your client rather than assume every setup needs an app. [Meta setup](). Meta's rules documentation says server rules are in limited availability. Rules can deny actions on particular ad accounts or product catalogs, and budget limits include percentage changes, absolute changes and ceilings. These are execution constraints; a separate queue for a person to approve each specific change was not verified in the reviewed documentation. [Rules and best practices](). The reviewed Meta pages do not establish a separate published server price, an exact beta launch date or a universal paused-creation default. We leave those unverified rather than repeat another vendor's description of Meta. Use our [official Meta server setup guide]() for client steps and [official-server limitations]() for the boundaries to investigate. ## Which options are hosted, and what can they change? Most commercial options below host the connection. GoMarble's open-source package is a distinct local implementation; a provider's hosted service and its repository need not expose the same tools. For app, token and maintenance ownership, see [what self-hosting a Meta Ads MCP server takes](). | Option | Deployment | Documented workflow | Important boundary | | --- | --- | --- | --- | | Meta's official Ads MCP | Meta-hosted | Reporting and ad management | Meta permissions and available server rules; verify each operation. | | AdKit | Hosted | Campaign, creative and account changes retained as drafts | Provider says approval is required before live changes. | | GoMarble open source | Local Python | Account objects, insights and activity retrieval | Reviewed README lists read tools; inspect installed code and token scopes. | | Markifact | Hosted | Reporting, campaign/creative management and connected marketing operations | Approval-based writes documented; exact queue boundary needs workflow verification. | | Pipeboard | Hosted; local installation also documented | Advertising reads, campaign creation, creative and budget changes | Explicit write confirmation and account/token scope. | | Birch MCP | Hosted | Inspect rules/logs; enable, disable and edit existing automation rules | Rule-management writes can affect later automation; not just read-only inspection. | | AdManage | Hosted | Creative launches, reporting, drafts, rules and edits | New connections read-only; explicit switch enables write tools. | | TheOptimizer | Platform connector | Agent access to its ad-automation engine | Detailed MCP tool scopes and per-change approvals not verified. | | Blend MCP | Hosted | Cross-platform reporting and edits | Session scoping and paused new campaigns documented; not per-change confirmation by default. | | NotFair | Hosted | Live reads and staged Meta mutations | Explicit human approval and operation provenance documented. | | Adspirer | Hosted | Multi-platform campaign preparation and management | New campaigns described as paused for review; exact review boundary needs verification. | | Ryze AI | Hosted | Meta insights, Graph reads, Graph writes and creative uploads | Published instructions use client tool approvals for write/delete/upload tools. | | Plainrouter | Hosted | Read Meta account evidence and signal health; pause or resume, change daily budgets, upload images and create paused ad copies | Ask holds every change. Full can run specified changes after policy checks; budget increases, campaign/ad-set creation, ad-set duplication and campaign/ad-set resumes still need a person. Executed changes get decision receipts. | ## What does each option cost? These are the providers' published prices or pricing boundaries checked October 29, 2026. They exclude ad spend, AI-client subscriptions and any infrastructure you operate. Promotional rates, billing intervals and limits can change; confirm the linked page and checkout before buying. | Option | Published entry price or model | Pricing source | | --- | --- | --- | | Meta's official Ads MCP | Separate server price not verified in reviewed documentation | [Meta overview]() | | AdKit | $29/month Single Project and $49/month Multiple Projects shown; 7-day trial | [AdKit pricing]() | | GoMarble open source | MIT-licensed code; you operate the runtime | [Repository]() | | Markifact | Pro from $79/month; credit allowances and execution logs vary | [Markifact pricing]() | | Pipeboard | Free: 30 weekly tool executions and 2 account slots; Pro $29.90/month | [Pipeboard pricing]() | | Birch | Subscription tied to connected ad spend; selected tier and MCP entitlement need confirmation | [Birch pricing]() | | AdManage | From £199.99/month; additional account allowances cost extra | [AdManage pricing]() | | TheOptimizer | Monthly Starter $199 up to $20,000 ad spend; overage and higher tiers apply | [TheOptimizer pricing]() | | Blend MCP | $29/month for 500 operations; $79 for 2,500; $199 for 10,000 | [Blend pricing]() | | NotFair | Free after trial: 300 operations/month; Growth lists $950/year and monthly billing | [NotFair pricing]() | | Adspirer | Free: 15 tasks/month; paid task allowances and billing intervals vary | [Adspirer pricing]() | | Ryze AI | Paid Ads $89/month listed, with a 7-day trial | [Ryze server reference]() | | Plainrouter | Free below €300 monthly observed ad spend; next band €19/month; MCP calls not metered | [Plainrouter pricing]() | ## AdKit: review drafts before publishing AdKit documents a hosted Meta workflow for campaign changes and creative work. Changes remain in AdKit as drafts until the user approves them, so its stated review boundary is before a live account change, not only before activating a newly created ad. Check the exact draft, account and operation in the reviewer view. [AdKit Meta MCP](). ## GoMarble open source: operate a local reporting server GoMarble's MIT-licensed Facebook/Meta Ads repository documents a local Python runtime with object reads, insights, pagination and activity history. Its installer can obtain a token through GoMarble; manual token configuration is also documented. This is the open-source package, not a claim about every GoMarble hosted product. The reviewed README's tool list does not establish campaign-write support; inspect the installed version before extending its access. [Official repository](). ## Markifact: connected marketing operations with approval-based writes Markifact documents Meta reporting and campaign, ad-set, audience and creative management alongside other marketing integrations. Its Meta page describes approval before writes go live. That establishes the advertised approval workflow, but we have not observed its authenticated reviewer interface or verified that every operation uses a separately persisted proposal queue. [Markifact Meta MCP](). ## Pipeboard: a family of advertising connectors Pipeboard's Meta server documentation covers hosted MCP access, with local installation as an advanced option. Its server family also covers Google, TikTok, Snap and Reddit. Documented Meta operations include launches, creative uploads and budget changes, with explicit write confirmation and paused new campaigns where supported. Confirm the account scope, selected token permissions and each platform's actual behavior. [Pipeboard Meta repository](). ## Birch MCP: inspect and change existing automation rules Birch's current tool directory includes read access to rules, execution logs, workspaces, account structure and Stage results, plus six rule-management write tools. These can enable or disable a rule, change its accounts or filters, rename it, move it or duplicate it. Enabling a rule can influence later account changes by the automation engine. A connector-specific human approval queue was not verified; inspect rule effects before granting those writes. [Birch MCP](). ## AdManage: read-only first, then optional writes and drafts AdManage documents new OAuth connections as read-only. Enabling read/write in MCP Access unlocks launches, uploads, edits, deletes, rules and automation changes. It also exposes draft-creation and draft-launch tools for review workflows; their presence does not mean every write is forced through a draft. Its guide explains workspace switching, so verify the active connection before changing an account. [AdManage MCP](). ## TheOptimizer: connect to an automation workspace TheOptimizer advertises an MCP connector into its campaign monitoring and automation engine. Its platform supports rule-driven pauses, status changes, budgets and other optimization tasks across connected channels. The reviewed public pages do not verify the connector's complete tool list or a per-operation approval contract. Obtain those details before treating platform automation capabilities as unrestricted MCP tools. [TheOptimizer](). ## Blend MCP: cross-platform edits with session scoping Blend documents hosted Meta reads and edits as part of a multi-platform MCP service. Its safety page describes session/account scoping and new campaigns starting paused, alongside platform change history. It also explains that its model is not a confirmation dialog on every action. Those controls do not establish mandatory review before editing an active object's budget; check that boundary separately. [Blend Meta MCP](), [Blend safety](). ## NotFair: staged writes and operation provenance NotFair's Meta connector uses hosted OAuth and exposes live account/reporting reads plus staged mutations such as pauses, budget changes and renames. Its Meta documentation says each write needs explicit human approval before reaching the Marketing API and records operation provenance. It explicitly does not claim its Google-style impact snapshots or undo for Meta. Verify where the exact proposed change is reviewed and stored in your chosen client. [NotFair Meta MCP](). ## Adspirer: campaign preparation across platforms Adspirer describes a hosted MCP product for existing AI clients, with multi-platform campaign preparation and management. Its public workflow shows new campaigns created paused for review and activation after approval. We have not observed the authenticated workflow or verified a separately persisted approval queue covering all edits to active campaigns. Check those boundaries against the job you want to delegate. [Adspirer](). ## Ryze AI: hosted Graph reads and writes Ryze AI at get-ryze.ai documents a remote OAuth MCP server with Meta reporting, Graph reads, Graph writes/deletes and image uploads, alongside other marketing integrations. Its setup instructs Claude users to keep write, delete and upload tools on the client's approval setting. Treat that as client tool confirmation; a separate provider-held review queue was not verified. Do not import its claims about Meta's official server into this comparison. [Ryze Meta MCP reference](). ## Where does AdAmigo fit? AdAmigo's own site describes an AI media buyer for Meta and Google, with review-first or automatic application of recommendations. An external MCP server endpoint and client setup were not verified on the provider pages reviewed, so we do not list it as a verified MCP connector above. It is an adjacent option if you want the provider's own agent rather than connect your existing client. [AdAmigo](). ## Plainrouter: choose approvals and your own conversion count Choose Plainrouter when you want your agent to work on Meta ads with a review queue and a record of what ran, while checking Meta’s figures against your own count. The hosted server reads account context, retained campaigns and metrics, the creative library and stored signal health. It supports pauses and resumes, daily-budget changes, image uploads and paused ad copies. See [Actions]() and the [AI-agent workflow]() for the supported tasks. Ask mode holds every change for a person. In Full, pauses, budget decreases, switching an individual ad on, image uploads and paused ad copies without a budget can run automatically after policy checks. Every budget increase waits for a person’s approval in both Ask and Full modes. Passing workspace policy does not waive that approval. Creating a campaign or ad set, duplicating an ad set and resuming a paused campaign or ad set also always need a person. Changes that break a rule are blocked and appear in the Inbox. Every change that runs gets a decision receipt. The kill switch stops further writes; it does not undo earlier ones. Plainrouter’s own arrivals and conversions sit beside Meta’s figures. Stored signal diagnostics help investigate delivery and matching; they do not measure incrementality or establish which ad caused a sale. You can use Plainrouter beside Meta’s official server, but a change sent through Meta’s server gets no Plainrouter approval or receipt. For production, sign up, connect Meta and start with Read access. The [MCP quickstart]() covers the connection. ## A real Plainrouter sandbox read: what it proves On October 11, 2026, we initialized an MCP session at ` https://plainrouter.com/mcp/sandbox `, listed its tools, then made this read request. This is synthetic sandbox data, not a production account or a competitor test. ```json {"jsonrpc":"2.0","id":3,"method":"tools/call","params":{"name":"get_account_state","arguments":{}}} ``` Selected fields from the actual structured response: ```json { "sandbox": true, "scope": { "ad_account_id": "act_SANDBOX", "workspace": "Sandbox Workspace", "destination": "SANDBOX_DATASET" }, "capabilities": { "signal_health": true, "verified_performance": true, "marketing_api_reads": false, "writes": false } } ``` The returned tool list contained ` get_account_state `, ` get_signal_health `, ` get_performance ` and ` validate_sandbox_event `. The sandbox exposes read and validation tools only and performs no writes. It demonstrates a working connection and synthetic response, not an approval flow: the proposals, policy checks and approval states described elsewhere on this page come from product documentation and were not demonstrated by this run. Conceptual product illustrations are not real execution output. ## A read-only Meta performance prompt Start with the account and the evidence available to your connection. This prompt uses Plainrouter's published read tools; it does not ask for an ad-level performance report that those Signals tools cannot provide. > Diagnose measurement for the account bound to this connection, without making changes. Call ` get_account_state `, then ` get_signal_health ` with ` days: 7 ` and ` get_performance ` with ` days: 7 `. Report the account, timezone, returned reporting windows and measurement timestamps. Separate completed-day arrivals, today's incomplete count, last-24-hour delivery diagnostics and stored Meta reconciliation. List observed discrepancies, unknowns and the next evidence needed. Do not infer ad-level CPA or creative fatigue from account totals. Do not call verification, proposal or write tools. The output should identify a specific evidence gap, such as accepted events without a confirmed destination result, or stored reporting that is too old for the question. Missing history is unknown, not zero. These tools read stored measurements; the requested window does not force a fresh Meta report. [Tool inputs and limits](). If measurement is the issue, use the [missing-conversion investigation prompt](). If the evidence points toward creative, use the [ad-fatigue review prompt]() with comparable ad-level data. Synthetic sandbox results only demonstrate connectivity. This is a documentation-based worked example, not an observed sandbox or production execution. ## Worked example: review a paused creative variant Suppose you have chosen an existing Meta ad and a replacement image already present in the connected account's creative library. Select Ask before submitting the proposal if you want to review it before execution. In Full, a policy-allowed proposal can queue automatically; asking the agent to stop after submission does not prevent that execution. This example follows Plainrouter's [published creative workflow](); it is not an instruction to activate an ad. | Step | What happens | What to check | | --- | --- | --- | | Read | ` get_account_state ` confirms scope; ` get-creative-library ` retrieves assets | Use the intended account, a human-selected source ad and an asset returned for that account | | Propose | ` duplicate-ad-with-creative ` creates a Plainrouter proposal | Review the source ad, replacement asset, fixed paused status, policy reasons and execution-mode snapshot; this call does not create a Meta ad | | Ask | A policy-allowed proposal waits for your approval | Approval is authorization, not proof that Meta accepted or completed the change | | Full | A paused ad-copy proposal without a budget can run automatically after policy checks | Budget increases, campaign or ad-set creation, ad-set duplication and campaign or ad-set resumes still need a person; rule violations are blocked in the Inbox | | Historical Suggest only | A pending historical approval records **Approved without execution** | No Meta copy is created; changing policy later does not make this existing proposal executable | | Verify executed work | The receipt identifies the new Meta ad for exact verification | Check the account, creative and paused state; pending verification is not a completed result | A new JPEG/PNG needs a separate ` upload-asset ` proposal first. Wait for its policy-allowed, verified upload, then read the creative library again for the returned image asset ID. A privately staged file is not yet a Meta image. Inbox also groups historical Suggest only approvals under **Landed**, so inspect the actual status and execution-mode snapshot before interpreting that label as a provider write. A verified paused copy remains paused and does not demonstrate better ad performance. Pause, resume, and budget changes can run through the governed pipeline; unsupported older action types never run. The kill switch stops further writes and never undoes a landed change. Review the [image specifications]() before preparing an asset. ## Compare approval behavior before enabling writes Three different controls can all be described as “approval”: an AI client asking to call a tool, a provider retaining a draft for review, and a policy authorizing execution. Establish which one your workflow actually uses. | Option | Documented change control | What to verify | | --- | --- | --- | | Meta Ads MCP | Permission-scoped tools; available rules deny actions and cap budget changes | Client confirmation, applicable rules and exact operation; human proposal queue not verified | | AdKit | Provider-retained drafts before live changes | Exact reviewer view, account and approved inputs | | GoMarble open source | Reviewed README lists retrieval tools | Installed version and token permissions before extending it | | Markifact | Approval-based write operations | Review location and persistence for the selected operation | | Pipeboard | Explicit write confirmation; paused new campaigns where supported | Client/server boundary and account/token scope | | Birch MCP | Reads plus write tools managing existing rules | Whether a changed or enabled rule can execute later; approval queue not verified | | AdManage | Read-only default; optional writes and draft workflows | Write switch and whether the selected operation actually waits in a draft | | TheOptimizer | Execution-capable platform connector advertised | Exact MCP scopes and per-change approval contract not verified | | Blend MCP | Session/account scoping and paused new campaigns; not every edit requires confirmation | Rules governing active-object edits and any added human review | | NotFair | Writes staged for explicit human approval; Meta operation provenance | Review location and exact mutation; do not assume Google undo applies | | Adspirer | New campaigns paused for review; approval of high-impact work described | Persisted review boundary and coverage of active-object edits not verified | | Ryze AI | Client approval settings for write, delete and upload tools | Keep approvals enabled; separate provider proposal queue not verified | | Plainrouter | Ask holds every change. Full can run pauses, budget decreases, individual-ad activation, image uploads and paused ad copies without a budget after policy checks. | A person still approves every budget increase, campaign/ad-set creation, ad-set duplication and campaign/ad-set resume. Check the decision receipt for each executed change; Meta-server writes do not receive one from Plainrouter. | ## A worksheet for choosing an MCP approval workflow **If you need to approve changes before execution, verify that the specific operation waits for review before it changes Meta.** A client confirmation, a connector-wide write switch and creation of a paused ad answer different questions. MCP recommends human oversight but does not mandate a particular user interface or a provider-side draft queue. [MCP tools specification](). Copy this worksheet for each shortlisted provider. Name one operation, such as replacing an image on a copy of an existing ad, and use the same requirement for every option. Record the provider, client, plan, date and evidence link. Label each finding **observed**, **documented** or **unknown**; a sales-page claim is not an observed result. | Check | Evidence to record | Decision it supports | | --- | --- | --- | | Account scope | Selected workspace/account and where access is restricted | Does the connection reach only the accounts you intend? | | Before approval | The proposed inputs and whether any Meta object already exists or changed | Does review happen before the provider write, or only before activation? | | Approval boundary | Who approves, where they review the exact change, and what happens if those inputs change | Is this a client tool prompt, a provider draft queue or both? | | Result evidence | Proposal/operation ID, execution status and resulting Meta object/state | Can you distinguish approved, executed and verified? | | Uncertain result | The documented way to look up a pending operation or resolve a timeout | Can you establish the outcome before retrying a possible write? | **Example interpretation:** a connector creates a paused ad and then asks whether to activate it. That offers review before activation; it does not meet a requirement that no Meta object be created before approval. A connector that retains a draft may meet that requirement, but you still need evidence for the selected operation. These are hypothetical examples, not findings about a named provider. Start with documentation and a read-only connection. Use a provider sandbox or dedicated test account for any authorized execution checks; an unavailable test remains unknown. Tool descriptions and annotations alone do not prove enforcement. The [MCP specification]() explicitly treats annotations from untrusted servers as untrusted. Choose the option with evidence for your required operation and review boundary. If that evidence is missing, keep the write workflow unevaluated rather than calling the provider unsafe or assuming the control exists. For Plainrouter, compare the [published creative workflow]() with the worked example above; synthetic connectivity results do not validate production execution. ## Frequently asked questions ### Is Meta Ads MCP the same as Facebook Ads MCP? Both names commonly describe an MCP connection to Meta advertising, including Facebook and Instagram. The provider still determines the tools, account permissions, and supported actions available through that connection. ### Do I need a paid MCP provider? Not necessarily. Start by comparing Meta's official access and local open-source options against your task. A hosted subscription may be worthwhile for managed setup, additional integrations, support, or a review workflow. Compare total setup and maintenance effort alongside the published price. ### Does MCP access mean an agent can spend money? The protocol name alone does not answer that. Look at the exposed tools, account permissions, and execution controls. Some connections expose retrieval tools; others can create or edit campaigns. A client confirmation prompt is not the same as a separate provider approval queue. ### Which option should I try first? Choose Plainrouter if you want an approval queue for Meta ad changes and an independent arrival and conversion count beside Meta’s figures. Use Meta’s official server when its own reporting and ad-management tools cover your workflow. Start production access by signing up in Plainrouter, connecting Meta and selecting Read access; add Write access when you want proposals. ### Which MCP clients can use these options? Client support depends on transport and authentication, not the provider name alone. Match remote HTTP/OAuth or bearer-header requirements to Claude, ChatGPT, Codex, Cursor or your chosen client; a local Python server requires a client or bridge that can launch it. Use each provider’s linked setup documentation rather than assume one config works everywhere. ### Should I choose a hosted or self-hosted Meta Ads MCP? Hosted services operate the endpoint and usually manage platform sign-in; you still own account access and change permissions. A local or self-hosted server gives you the runtime and code to inspect, with responsibility for credentials, dependencies and maintenance. GoMarble’s repository and hosted products are separate choices.