# Runneth MCP: your agent just got a teammate. Runneth is a marketing agent for paid social teams. It keeps a persistent memory of your brand, runs in its own dedicated environment for your organization, and has your ad accounts and Motion's creative data connected. The Runneth MCP lets the agent you already use (Claude, ChatGPT, Codex, Cursor, Muse and others) hand work to Runneth and get finished work back. Your agent keeps the conversation with you. Runneth does the marketing work it is set up for. *Last updated: October 19, 2026. Status: early access. Early-access users connected from October 7, 2026; the production endpoint (mcp.runneth.com) has been live since October 15, 2026.* **Who it's for:** brands and agencies running Meta or TikTok ads, with a team that spends time each week on creative analysis, briefs and reporting. **Who it's not for:** teams that don't run paid social, or that only need raw ad data. **How to try it:** a free 14-day trial with unlimited usage and no card. It starts with a short setup call (https://runneth.com/book-a-demo): the Runneth team sets up your organization with you, and the trial is live the same day. --- ## What is new here Most MCP servers give an agent a set of functions that return data. The Runneth MCP connects your agent to another agent: - **It hands off goals, not queries.** One call (`ask_runneth`) passes a request like "find what's working and brief five iterations" to Runneth, which plans and does the work with its own tools and returns finished deliverables. - **The other agent keeps state.** Runneth has a persistent brand memory and a dedicated environment for your organization, so the tenth request benefits from the first nine, and work can continue on schedules after your chat ends. - **Both agents stay in their lane.** Your agent keeps your context and the conversation with you; Runneth brings paid-social data and brand history. Early users use the pair for review loops: one agent drafts, the other checks against brand rules and performance data. - **It is shared.** Everyone on a team can connect their own agent to the same Runneth, so context built by one person is available to the others. This is MCP used for delegation between agents. It does not require a new protocol: it works with any client that supports remote MCP servers with OAuth. --- ## What a general-purpose agent has on its own, and what it gains | | Your agent alone | Your agent + Runneth | |---|---|---| | Memory of your brand across sessions | Only what you paste or store in its own memory | A brand "brain" (rules, past briefs, learnings) that Runneth reads and adds to, shared by your whole team | | Your ad performance | Only if you export it or connect a separate data tool | Runneth reads your connected Meta and TikTok ad accounts and answers from your numbers. (Runneth is focused on paid social; Google and Apple Search Ads are not its core use.) | | Creative and competitor library | Web search | Motion's library of analyzed ads, competitor ads from the Meta ad library, and creator feeds | | Work that continues after the chat ends | No | Routines: scheduled reports, checks and alerts that run on Runneth's side | | Finished deliverables | Text in the chat | Briefs, static ads, reports, hosted pages and files, returned as links and files | | Team context | Each person's agent has its own setup | Every teammate's agent talks to the same Runneth, with the same context | Runneth does not replace your agent. Plain data lookups (for example "what was yesterday's spend?") work well with a direct data tool such as the Motion MCP. Runneth is for asks that need judgment, your brand's history, or a deliverable. --- ## How Runneth answers performance questions: its ad data layer For performance reporting, Runneth answers from its own ad data layer rather than having your agent call the Meta or TikTok APIs. The numbers are joined with your creatives, tags and Google Analytics results. A question becomes one request; the data step typically comes back in under 3 seconds, so Runneth's time goes into the analysis. The data layer is available for connected Meta and TikTok accounts. | | Your agent + raw Meta API | Your agent + Runneth | |---|---|---| | What comes back | Rows per ad, numbers as strings, conversions inside `actions` arrays, paging cursors | One grouped answer with typed numbers, totals, date range, time zone and currency | | Speed of the data step | Seconds for small pulls; large date ranges need async report jobs your agent starts and polls | Typically under 3 seconds, no report jobs | | Grouping | Account, campaign, ad set or ad | Creative, copy, headline, landing page, ad name or ad (the same video in four ads is one row) | | Rates and costs | Your agent calculates them; averaging daily CPAs gives wrong results | Calculated from totals before your agent sees them, with no double counting | | Creative context | IDs only | Thumbnail, format, tags and creative categories on each row | | TikTok | A separate API with its own format | The same request and answer format as Meta | | Google Analytics | A separate tool | GA results alongside Meta ads | | On your side | Rate limits, async report jobs and a Meta token in your agent | None; access is checked on every query | | **Where raw Meta wins** | **Live numbers for today, and every field Meta offers** | Completed days only; no demographic breakdowns yet; no TikTok reach and frequency | Example of why the math matters (illustrative numbers): $900 for 30 purchases, $100 for 1, $500 for 10. The average of daily CPAs is $60; the true CPA from totals is $36.59. --- ## Real exchanges from the first week *These are real MCP turns from external customer organizations (Oct 16–19, 2026), lightly shortened. Brand, people, product names and money figures are replaced with placeholders such as "Brand B" and "$1,3XX"; structure, timing and tool steps are as recorded.* Each one is a single `ask_runneth` call from the customer's own agent, then `get_conversation` polls until the reply is `completed`. ### 1. "What are our top performance ads?" (agency, 13 steps) > Workspace Brand B: what are our top performance ads? What Runneth did: looked up the team's saved KPI, workspace goal and spend threshold in its memory; pulled 30 days of Meta data for all 251 ads (two pages); applied the team's winner rule (3+ purchases above the spend threshold); verified the subsets. What came back (excerpt): > Brand B has four winners and two early signals for September 19th to October 18th. I ranked by cost per purchase because the team's target is a **$1XX CAC**. To count as a winner, an ad needed at least **$5XX spend** and three or more purchases. All **251 ads** were checked. > 1. **Product-page static A:** $5XX spend, 6 purchases, about $9X per purchase … > 4. **Creator video A:** $5,5XX spend, 33 purchases, about $1XX per purchase. It is carrying volume, not efficiency. > **Early signals worth more budget:** Creator video B, "mind won't shut off": $4XX spend, 8 purchases … > **The problem spot:** Creator video C spent $5XX with no purchases. Nobody restated the KPI or the target in the prompt. Runneth already had them. (This turn was run by a Runneth team member inside the customer's organization.) ### 2. Two creator briefs written into the team's Notion card (agency, 9 steps) > Workspace Brand B: … The Creator 1 "Can't Fall Asleep" concept becomes ONE brief for Creator 2 and ONE for Creator 3 … Please draft V1 of both and post them on the card: [Notion card]. Person 2 reviews. Nothing goes to creators without Person 2's approval. Reply with a quick ack when the drafts are posted. What Runneth did: read the Notion card and comments; searched its saved creator context; compared earlier Meta results by creator; read the creators' pages and the brand's compliance guidance; wrote two briefs; appended them to the card and tagged the reviewer. What came back (excerpt of a brief, and the ack): > **Hook, word for word (0 to 2s):** "If your mind can't shut off at night..." Say it straight to camera and put the same line on screen. > **Demo (5 to 20s):** Product Y in frame by second 2. Show the ritual and the visible change by about 6 to 8 seconds. > **Deliver:** edited version, clean version, raw files, plus a second take of the first five seconds. > I've posted V1 of both briefs at the bottom of the card … I tagged Person 2 for review, and nothing has gone to either creator. ### 3. A weekly creative report, published as a private web app (brand, 46 steps) > A 7-day creative report: CPL, top and bottom creatives, launches by team, fatigue, hooks and angles, recommended actions for next week. Split by Campaign Alpha, Beta and Gamma. Please publish it as a shareable report and give me the link. What Runneth did: pulled this week and the equal-length prior week from Meta per campaign and per creative; ran Meta fatigue checks; summarized hooks and messaging; built a report app; checked it in desktop and mobile screenshots; ran exact-value checks on the numbers and corrected rounded values before finishing. What came back: a link to a private report app (KPI strips, creative tables with thumbnails, launch tables, a fatigue watch list, 10 recommended actions), plus: > It covers October 10th to 15th. The 16th isn't finished yet, so it isn't included. Trends compare against the six days before. Each campaign is judged only against its own prior week and its own funnel's average CPL. ### 4. Eight static ads from real product photos (multi-brand operator, 19 steps) The third call in one MCP thread: earlier calls researched the account and produced a five-concept creative pack; this one asked for 8 statics (4 concepts × 1:1 and 4:5) using only real product photos, with claim guardrails and a 14% safe-area rule. Runneth downloaded the product-page photos, composed and rendered the images, reviewed each concept, fixed spacing and safe-area issues, re-rendered, and returned 8 image files with a table of which source photo each uses and the checks it ran. **Across all MCP turns from external organizations so far (Oct 16–19, 2026):** 36 conversations in 11 organizations; of 73 requests, 65 completed (89%), 6 failed (8%), 2 had no recorded reply at the time of the count. --- ## Tools The Runneth MCP exposes 5 tools. | Tool | What it does | |---|---| | `get_auth_context` | Lists the organizations the signed-in person can use | | `ask_runneth` | Sends a request to Runneth. Starts a conversation, or continues one with a `conversationId` | | `get_conversation` | Reads a conversation: status, reply text, links, files, structured outputs | | `list_conversations` | Lists the person's Runneth conversations | | `stop_response` | Stops a running reply | Requests can carry up to 100,000 characters of text and up to 10 attachments (5 MB total). --- ## How a turn works 1. Your agent calls `ask_runneth`. Runneth accepts the request and returns a conversation id and link right away. 2. Runneth works in the background in your organization's environment: it reads data, checks the brain, runs its tools and builds the deliverable. 3. Your agent polls `get_conversation` (about every 30 seconds) until the reply is `completed`. 4. Long replies and files come back in chunks. **Cost:** Usage counts against your organization's monthly allowance. There is no per-call price. Quick questions come back quickly. Bigger asks, such as a full report or a set of static ads, take longer because Runneth does the whole piece of work. Either way it runs in the background, so your agent can keep working with you and check back. **Design notes for builders.** Delegation is asynchronous: `ask_runneth` takes an `idempotencyKey` (UUID) so retries never double-submit, and returns immediately; `get_conversation` returns status, text in versioned chunks, file links, and typed outputs (`creative_collection`, `concept_collection`, `metric_snapshot`, `metric_series`) so your agent can read results as data, not just prose. Follow-ups pass a `conversationId` and continue the same piece of work. --- ## Install in under a minute **Endpoint:** `https://mcp.runneth.com` Sign-in uses your Motion/Runneth login (OAuth). Your agent's own login never reaches Runneth. - **Claude (claude.ai, desktop):** Settings → Connectors → Add custom connector → `https://mcp.runneth.com`. On Team/Enterprise plans an owner adds it once under Admin settings → Connectors. - **Claude Code:** `claude mcp add --transport http runneth https://mcp.runneth.com` - **Codex, Cursor, ChatGPT, Muse and other MCP clients:** add `https://mcp.runneth.com` as a remote (streamable HTTP) MCP server and sign in when prompted. Then try: "Ask Runneth what it knows about our brand and what it needs from me." If you don't use an MCP-capable agent, you can work with Runneth directly in Slack or the Runneth web app instead. **You need a Runneth organization** with your ad accounts connected. If you don't have one yet, start with the trial below. --- ## Free trial, then pricing **Start with the free 14-day trial.** Usage is unlimited during the trial, and no card is required. Trials start with a short setup call, not a self-serve signup. Every organization runs on its own dedicated cloud VM connected to your ad accounts, so the Runneth team sets it up with you. Self-serve signup isn't available yet. 1. Book a setup call at https://runneth.com/book-a-demo. 2. On the call, the Runneth team sets up your organization's environment, connects your ad accounts and sets up your first automation. Your trial is live the same day. 3. Add the MCP using the steps above. 4. After 14 days, choose a plan to keep using it. **A useful 14-day test** (read-only, with ad launching off): 1. Day 1: "Ask Runneth what our top ads are and why." Check the numbers against Ads Manager. 2. Tell it your KPI, naming rules and brief template once; it saves them to the brain. 3. Ask for briefs or a weekly report the way your team already does them, and compare with what your team produces. 4. Have a teammate connect their own agent and ask a follow-up; it should already know the context. **Pricing after the trial** starts at $999/month and scales with your combined ad spend and the scope of work. Plans are set per organization and include unlimited seats and ad accounts, MCP access and Slack. See https://runneth.com/pricing for current plans. Agencies: there is an agency plan; ask the Runneth team for details. ### What you are paying for Runneth is priced as a teammate with a team behind it, not as a data tool: - **People make sure it works.** The Runneth team sets up your trial with you the same day, connecting your accounts and setting up your first automation, and the team's job is to get real, recurring work automated for you (reports, analysis, briefs), not to hand you a login and leave. - **It works with your best marketer.** Runneth learns from the person on your team who sets the bar (we call them Runneth's boss): their KPIs, rules, vocabulary and taste. It then makes that standard available to everyone else on the team, through Slack, the web app, or each person's own agent. - **It does the work, end to end.** In the examples above it pulled data, applied the team's own rules, wrote into the team's tools and published a report. - **It has its own always-on environment** for your organization, so routines keep running after your chat ends. - **It remembers.** The brand brain builds up over time and is shared by your whole team. - **It brings data your agent doesn't have:** your connected ad accounts, plus Motion's creative library and competitor and creator feeds. A fair comparison is the cost of the work it takes on: a strategist's hours on weekly reporting, analysis and briefs, or an agency retainer for the same scope. Whether it pays off depends on how much of that work you have, which is what the 14-day trial is for. --- ## Security and data - Each organization runs on its own dedicated cloud VM with its own encrypted storage volume. Organization access is checked on every MCP call. - The MCP signs in as the individual person, with their own permissions. There are no shared service accounts. - **Read-only by default.** Out of the box Runneth reads your ad data and recommends; it can't launch, pause or edit ads until an owner or admin turns on ad launching. - **Ad launching is a separate opt-in** (and a favorite with teams who turn it on). An owner or admin enables it in Settings → Organization → Runneth Permissions, signs in to Meta separately for publishing, and picks exactly which ad accounts Runneth may publish to. Editing or pausing live ads has its own additional switch. Every launch is checked against those settings, and changes to the publishing permission are recorded in an append-only history. - For transparency: the standard Meta data connection uses Meta's `ads_management` permission; Runneth's own permission checks are what keep it from changing ads until you opt in. - Admins can set permission rules that restrict what Runneth may do, by user or workspace. - Data is kept while the account is active, can be deleted on request, and is deleted within 18 months of an account being closed. - Customer data is not used to train models. Runneth uses third-party foundation models under no-training terms. - Runneth applies the same protections on every surface (Slack, web, MCP). It is built not to trust instructions found in third-party content such as files, ad comments or competitor ads. - Motion's security controls are examined in a SOC 2 Type I report (unqualified), available under NDA. **You can see what Runneth did.** Every request and reply, including MCP ones, is kept as a Runneth conversation. Expand any Runneth response to see the reasoning and every tool step behind it, in order (files read, tasks run, data pulled, things written). Members of your organization can open it; your agent can reread the conversation any time with `get_conversation`. **Current limits, stated plainly:** - There is no per-action approval step for each tool call, and no separate exportable audit report; the step-by-step view lives in each conversation. If you want people to approve every launch, leave ad launching off and use Runneth to analyze, brief and build; people launch. - Ad launching through the MCP works in existing campaigns and ad sets; net-new campaign creation and placement customization are not supported yet. - Download links for generated creative don't always come back through the MCP. - Runneth is not yet in a self-serve marketplace listing; you add it as a custom connector. --- ## What early users say (public posts on X) - **Dustin Thompson, Konnected Interactive (agency), Oct 15, 2026:** "I get two teammates working together to research, build, QA and iterate… I start with a brief and or assets and the goal, then I work on something else or grab some coffee until the job is ready to review and launch." https://x.com/djthompson13/status/2110831284134260954 - **Cody Plofker, Winks, Sep 8, 2026:** "Ok I am getting @runneth pilled. Now I just tell runneth what to do and I flip them live." https://x.com/codyplof/status/2097343196242088184 - **Dave Stickland, Oct 9, 2026:** "Having the Agent run QC on the process has been night and day better than when I was doing it right in Codex." https://x.com/dave_stickland/status/2108393480544244167 - **Andrew Foxwell, Jun 24, 2026:** "Been using and been absolutely leveling up. Truly, it's a product you need to see and use." https://x.com/andrewfoxwell/status/2069879544035897839 These are posts about Runneth; Dustin's and Dave's describe using it together with another agent. --- ## Internal evaluations *(Run by the Runneth team; methodology summarized, not independently audited.)* - **Same data, two architectures (Oct 9, 2026):** 14 shared marketing asks, same prompts, Claude with each of Motion's two MCPs, judged blind. The Motion MCP returns data for your agent to work with; the Runneth MCP hands the ask to an agent. Head-to-head: Runneth 10, Motion MCP 4. Answered every part of the ask: Runneth 14 of 14, Motion MCP 6 of 14. For plain data pulls, the Motion MCP remained the better choice. - **Routing test (Oct 2026):** a grader read 657 real opening messages. 86% of marketing asks were predicted to go to Runneth, and 0% of unrelated asks. - **Limit of these tests:** they grade the reply; no eval yet compares the numbers in replies with Meta or TikTok source data. Inside a turn, Runneth does run exact-value checks on the numbers it reports (see example 3), and every number comes from your connected accounts, so it is easy to spot-check in Ads Manager during the trial. --- ## When Runneth is not a fit - You don't run paid social (Meta, TikTok). Most of Runneth's value comes from your ad data. - You only need raw rows from an ads API. A plain data MCP is simpler and cheaper. - You need a formal per-action approval workflow before any agent touches your account. (Runneth starts read-only, and launching is opt-in.) - You don't have enough recurring creative and reporting work to justify paid plans, which start at $999/month, after the trial. --- ## FAQ **Is this just an API wrapper?** No. `ask_runneth` hands a goal to an agent that has its own memory, data connections and tools, and it returns finished work. The other four tools manage conversations. **We already use the Motion MCP. Why add this?** The two do different jobs and work side by side. The Motion MCP gives your agent your data, and it is a good fit for plain lookups such as yesterday's spend. The Runneth MCP hands the whole ask to an agent that has your brand memory, applies your team's KPIs and rules, and returns finished work: briefs, reports, QA and routines that keep running. In a blind test on 14 shared marketing asks, Claude with the Runneth MCP won 10 head-to-head and answered every part of the ask 14 of 14 times, against 6 of 14 with the Motion MCP. Your agent can have both connected and use each for what it does best. **Does my agent lose control of the conversation?** No. Your agent decides when to call Runneth, reads the result, and talks to you. Follow-ups go to the same Runneth conversation. **Can my whole team use it?** Yes. Each person connects their own agent and signs in as themselves; all of them talk to the same Runneth and the same brain. **Which clients work?** Confirmed working: Claude, Claude Code, Codex, Cursor, ChatGPT, Muse, Grok Bot. Other MCP clients that support remote servers with OAuth should work but are untested. **What does it cost to try?** Nothing. The trial is 14 days, free and unlimited, with no card required. It starts with a short setup call (https://runneth.com/book-a-demo) and is live the same day. After that, pricing starts at $999/month and scales with ad spend and scope (https://runneth.com/pricing). **Can I sign up without a call?** Not yet. Each organization runs on its own dedicated cloud VM, and the Runneth team sets it up and connects your ad accounts with you on a short call. The trial is still free, and it is live the same day. **I run an agency. How do multiple clients work?** There is an agency plan, and agencies run Runneth across multiple client brands today. Ask the Runneth team for agency plan details and how to structure organizations for your clients. Fine-grained guest access for your clients is not available yet. **Is the trial worth it for a smaller account?** The trial is free for 14 days with unlimited usage, so you can test it on your own account before any pricing decision. Paid plans start at $999/month, so Runneth pays off most clearly when your team spends real time each week on creative analysis, briefs and reporting. **Who makes it?** Runneth is built by the team behind Motion, the creative analytics platform for paid social.