A multi-agent harness plans your marketing and executes it across SEO, analytics, ads and publisher platforms — Search Console, GA4, Google Ads, DV360, CM360, Ad Manager, AdSense, AdMob and BigQuery — then connects to 250+ more tools via Composio and your own MCP. Every write waits for your go-ahead.
Most marketing AI is a black box — you flip a switch and hope for the best. Agent Planners reasons through your goal in the open, adapts to any platform or model, and waits for your yes before anything changes.
Every action arrives as a plan you read before it runs.
Nothing changes on your account without your sign-off.
Every mutation is logged — who, when and why.
The agent reads your accounts and your own data, then answers with the next move — sourced, quantified and ready to approve.
Which campaigns are wasting budget this week, and what should I do?
3 Search campaigns are under 1.5× ROAS — Brand_Generic, Q3_Prospecting and Retarget_Broad — together burning ≈$3,420 / week.
Biggest leak: 14 search terms with spend and zero conversions (e.g. “free …”, “jobs …”).
Recommended: pause Retarget_Broad, add the 14 negatives, and shift 20% of budget to your 3.1× ROAS winners.
Every write waits for your approval before it ships.
Natively connects to Google's advertising, publisher & analytics platforms — plus 250+ more apps via Composio or your own MCP server. Powerful research, web & creative tools are built in, and every agent runs on the model you choose through one built-in AI Gateway.
A multi-agent system, an observable workflow engine, a real tool harness, schedules, teams, and multi-account execution — built to run your accounts, not just chat about them.
An orchestrator plus specialist agents — analytics, optimization, creative, research, safety and report — ship in the box. Each one runs on whatever model you choose, switchable any time on the Agents panel through one gateway.
The agent auto-clarifies the goal, judges what's needed, plans the steps, executes them, and emits a plan you can read. Plan → Execute → Summarize → Verify → Notify is generated for you — and when a change needs your call, it pauses and asks for approval.
Mid-run the agent picks the right tool, skill, MCP server and sandbox on its own — web search, Google Trends, Firecrawl, Exa neural search, a hosted browser, image/video generation and a Python sandbox — and manages its own context and memory. It reads its execution history and the account, self-repairs errors, and keeps the task on track.
Turn any task into a scheduled job — hourly, daily, weekly, or a raw cron in any timezone. Each fire re-evaluates against live data, so a recurring optimization never re-proposes yesterday's actions, and every run can email a summary.
Create multiple organizations and set each member's permissions — admin, editor, or member. Manage your skills, integrations and assets per org, with cross-org sharing when you want it and isolation when you don't.
The Context Picker runs across many accounts at once and pulls every connected platform into a single execution. Native support for Google Ads, DV360, Campaign Manager 360, Google Ad Manager, AdSense, AdMob, GA4, Search Console and BigQuery — plus 250+ more apps through Composio and your own tools via custom MCP.
The capabilities that make it operate like an experienced ad-ops engineer: it delegates, picks its own tools, asks before it acts, remembers, learns, and fixes its own mistakes.
An orchestrator decomposes the goal and hands steps to six specialist agents — analytics, optimization, creative, research, safety and report — each on the model you pick.
Mid-run the agent picks the right tool itself — your ad & analytics platforms, web search, Google Trends, Firecrawl, Exa neural search, a hosted browser, image/video generation, a Python sandbox, plus Composio and your own MCP servers.
Before it plans, the agent asks about anything ambiguous — accounts, thresholds, the real goal — so it never guesses with your budget.
Flip on read-only autonomy or auto-approve for hands-off runs and 24/7 schedules — or keep the approval gate on every write. Your call.
Per-account memory — target CPA/ROAS, protected campaigns, excluded geos, rules — is recalled on every run, so it operates the way you do.
It reads the account's change history and its own past runs to stay on track and avoid re-proposing yesterday's actions.
A successful run is distilled into a reusable Skill, so the agent follows your proven method next time instead of starting from zero.
It reads tool errors from results and self-repairs mid-run — re-validating, correcting arguments and retrying — instead of failing the task.
Turn a run into a client-ready slide deck or report — the agent drafts the narrative, charts and slides (built-in PPT generation), and you can edit them.
An independent critic reviews the draft plan before it runs — catching missing steps and known pitfalls from past runs, and revising it before anything executes.
When a step reveals more work than expected, the agent expands the plan on the fly — adding the steps it needs instead of stopping short.
If it hits a fork it can't resolve during execution, it pauses and asks you — in-app or by email — then continues with your answer.
Point agent LLM calls at your own OpenAI-compatible endpoint — Azure AI, vLLM, OpenRouter, or a model you host — instead of the built-in AI Gateway. LLM tokens on it are zero-rated, so you pay your provider directly, not platform credits.
Azure AI Foundry, vLLM, OpenRouter, or a model you host yourself — paste the base URL, model id and key.
Saving runs a live call against your endpoint first. A dead URL, key or model never persists.
Tokens on your endpoint cost $0 platform credits — you pay your provider directly. Tool, browser, sandbox and media calls still meter.
The problems it actually solves — and why teams don't switch back.
Most marketing AI is a black box — you flip a switch and hope for the best. Here every mutation pauses at a gate — in-app or magic-link email — and a single reject skips that step instead of aborting the run.
Join Ads cost with your own BigQuery / CRM mart and let pause and budget decisions run off the numbers you believe — not Google's conversion guesswork.
A Skills library plus per-account memory — target CPA, protected campaigns, rules, history — captures how you operate and applies it on every run.
One goal fans out across an entire MCC with a dryRun blast-radius preview first, v22 GAQL guardrails, sanity-checks, stale-state re-validation, and partial-failure tolerance.
Autonomy without visibility is a liability. Every run is observable end to end, every write is approval-gated, and every change is logged — so you get the speed of an agent with the accountability of a human-run process.
Every task is a step-by-step plan you can read — the reasoning, the exact tool and arguments — not a black box. Reads run in the open; nothing changes on your account unseen.
Plan → Approve → ExecuteEach mutation pauses until you say yes — in-app or by a one-time email magic-link, with the full reasoning attached. One reject skips that step; the run keeps going.
In-app or emailEvery mutation is recorded — who, when, the channel, the decision, and before/after detail. Nothing changes silently, and you can trace any change back to the run that made it.
Who · when · whyBefore a write executes, live account state is re-validated — a change planned against old data is rejected instead of applied blindly. Tool errors are read from results and self-repaired mid-run instead of silently failing the task.
Re-validated · self-healingOAuth tokens and connector keys (Composio, custom MCP) are encrypted at rest and scoped to your organization — never shared across orgs, never returned to the browser.
Encrypted at restPer-account memory enforces protected campaigns, target CPA/ROAS and excluded geos on every run. Organizations set admin/editor/member roles, so access matches who should actually approve what.
Roles · protected campaignsOne loop, end to end. Watch the three stages, or click through them.
Concrete, multi-step work across your whole marketing stack — SEO, analytics, paid ads, publisher revenue and whole-MCC scale — plus the CRM, Slack and reporting workflows that turn a marketing insight into business impact. Each from a plain-language goal, with every write gated by your approval.
Cross-source reads that turn raw account data into a decision: what moved, why, and what to do next.
Summarize the last 7 days — cost, conversions, CPA, ROAS, CTR — flag any campaign whose CPA moved >25% and give me 3 ranked next actions.
Find ads that are DISAPPROVED or SITE_SUSPENDED with their policy topics and email me the full list.
Summarize the last 7 days of change events grouped by user and resource; highlight budget changes over 20%.
Pause the leaks, mine wasteful search terms into negatives, and shift budget toward your winners — every write waits for your approval.
Pause campaigns under 1.5× ROAS across all accounts.
Flag search terms with cost over $20 and zero conversions, then propose campaign negative keywords.
Move 20% of budget from campaigns under 1.5× to my top 3 above 3×; show a before/after table.
From a brand Search campaign to a full Performance Max with generated assets, or a remarketing audience wired from GA4 — drafted paused, for your review.
Create a paused brand Search campaign with one ad group, 3 phrase keywords and a responsive search ad.
Create a paused PMax campaign with an asset group; generate any missing images and link them.
Read my GA4 behavior data, build a matching GA audience, and wire it into a remarketing campaign.
Fan one task across every account with batch operations, run bid experiments, and research the market — at agency scale.
Across the selected accounts, pause campaigns matching ‘Test_’ or ‘Legacy_’ but never anything with ‘brand’ — dry-run first.
Recommend device, geo and ad-schedule bid modifiers from 30-day data and run a 50/50 experiment for 14 days.
Study competitor sites and the market, then craft a launch strategy — keywords, angles and budget.
Marketing isn't just paid: monitor SEO in Search Console, analyze GA4, and optimize publisher revenue across AdSense and Google Ad Manager — all from the same agent.
From Search Console, find queries and pages that lost clicks, impressions or average position week-over-week and rank the biggest SEO drops.
Surface high-impression, low-CTR queries and page-2 keywords in Search Console, and propose the pages and titles to improve.
Break down conversions by channel and landing page in GA4, flag the biggest funnel drop-offs, and recommend where to invest.
Analyze AdSense earnings, RPM and top pages, flag drops, and recommend ad units and placements to test.
Review Google Ad Manager delivery — fill rate, impressions and eCPM by ad unit — and flag underperforming inventory.
Wire outcomes into the tools your business actually runs on — CRM, Slack, email, Notion — so a campaign insight becomes a sales lead, a real-time alert or a brief, instead of a copy-paste job.
When a visitor crosses my GA4 MQL threshold, create or update the contact in HubSpot with their campaign, source and last-touch channel.
Post an alert to #paid-media whenever daily spend crosses 120% of pace, or a campaign's CPA jumps over 30% day-over-day.
Every Monday, summarize GA4, Search Console, Google Ads and AdSense performance into one report and email it to leadership.
Research this month's top 5 competitors' positioning and ad messaging, and add a structured brief to our Notion workspace.
The reason people hand their accounts to an agent here: they can see every action before it runs, approve what writes, schedule the routine, and share one workspace with their team.
“What sold us is that every change shows up as a plan we read before anything happens — pause this campaign, shift that budget, add these negatives. Nothing moves without our click. We finally let AI into our Google Ads without losing sleep, and it cut spend on broad, off-brand terms in the first two weeks.”
Wary of letting AI touch live ad accounts — needed to see exactly what it would change first.
“It runs our weekly performance review on a schedule and drafts the cuts — low-ROAS pauses, wasteful search terms turned into negatives — every Monday before I'm even online. I approve straight from the email. It's like an analyst who never forgets and always shows their work.”
Too many campaigns to review by hand; the Monday reporting-and-cleanup grind ate half a day.
“We run it across the whole MCC with the team in one workspace — roles per member, shared skills and memory, and an audit trail of every approved action. Onboarding a new buyer is now sharing the org, not re-explaining each client account from scratch.”
A team of buyers managing dozens of client accounts, with no shared context or change history.
Representative scenarios reflecting how teams across ad-ops roles use the platform — not attributed to named customers.
Yes — security is our first priority. Every write pauses at an approval gate, every action lands in an operation / audit log, and you get email notifications for approvals and completions. Every integration, key and permission can be deleted and revoked at any time.
No. Reads are free; every mutation waits for a human — approve in-app, by magic-link email, or pre-approve a task when you explicitly want it hands-off. A single rejected step is skipped, not the whole run.
Both. Each pending write shows its title, the agent's reasoning, the exact tool and arguments, and the task goal. Approve it in the Workspace, or click a one-time magic-link sent to an approver's email. Before a write runs, the Safety agent re-validates live account state, so a plan built on stale data is rejected rather than applied.
Natively, over OAuth or a pasted token: Google Ads (deep, pinned to API v22, including MCC), Display & Video 360, Campaign Manager 360, Google Ad Manager, AdSense, AdMob, Google Analytics 4, Search Console, BigQuery, Meta Ads, TikTok Ads and Shopify. See platform-by-platform use cases on the Platforms page. Plus 250+ more apps (Gmail, Outlook, Slack, HubSpot, Salesforce, Notion…) via Composio and your own MCP servers.
Yes. Connect Google Ads with its MCC hierarchy and select many accounts in the context picker — the agent fans one goal across them. Cross-account writes always show a dry-run blast-radius preview first, then re-validate live state before applying.
Yes. Add a custom MCP (Model Context Protocol) server on the Integrations page — give it a name, a Streamable-HTTP URL and an optional token — and the agent auto-discovers and calls its tools. For pre-built SaaS actions (Slack, HubSpot, Salesforce and more) connect Composio instead.
Yes — the opposite direction from custom MCP servers above: here, an external agent calls in to run tasks on your accounts. Agent Planners exposes a REST API and an MCP server, both authenticated with per-organization API keys. Point Claude Desktop or Claude Code at the MCP server, or call the REST API from Codex, a script or a CI job — either way it runs on the same orchestrator, with the same approval gates and credit metering as a task started in the Workspace.
Yes. The orchestrator and each specialist agent run on whatever model you choose, switchable any time on the Agents panel through the built-in AI Gateway — GPT-5.5, Claude, Gemini, Qwen, DeepSeek and more. By default the orchestrator uses a quality-first planner and specialists use a fast, low-cost model.
Yes — Bring Your Own Model, included on any paid plan. Point every agent at your own OpenAI-compatible endpoint (Azure AI, vLLM, OpenRouter, self-hosted) and its LLM tokens are zero-rated — you pay your provider directly instead of platform credits. Tool, browser and sandbox calls still meter as usual, and saving runs a live test call first, so a dead endpoint never persists.
Usage is metered in credits, billed per organization. A credit is a fixed slice of real model and tool cost: tasks spend credits on language-model tokens plus any external tools they call (web search, image/video generation, browser sessions, the Python sandbox). Cheaper models cost far fewer credits, so most of your burn depends on which models your agents use — a typical optimization task runs roughly 400–800 credits.
Every organization gets free monthly credits (2,500 on the free tier) with no credit card required to start. Paid plans raise the monthly allotment and one-time top-up packs add credits that stack on your balance. See the Billing page for current pricing and to upgrade or top up.
Plan credits reset each billing period and don't roll over. One-time top-up packs stack on your balance and do not reset, so they're there until you use them.
Runs are blocked at zero credits until the next monthly reset, a plan upgrade, or a top-up pack — nothing is charged silently. You can watch your balance and usage (by day and by model) on the Billing page.
Yes. Turn any goal into a scheduled job on a cron — hourly, daily, weekly or a raw expression — in your timezone. Each fire regenerates a fresh plan against live data, so a recurring optimization never re-proposes yesterday's actions. You can set an approver email, enable auto-approve, or trigger a Run now any time.
Yes. Each connected account carries durable memory — target CPA/ROAS, protected campaigns (e.g. never pause anything matching "brand*"), excluded geos, pacing and custom rules — that the agent reads on every run and you can edit on the Memory page. You can also capture your methods as reusable Skills.
Yes. Every workspace is an organization with isolated accounts, integrations, skills, assets, audit log and billing. Create as many orgs as you need, set per-member roles (admin / editor / member), and fan one goal out across an entire MCC.
OAuth credentials and tokens are encrypted at rest and isolated per organization, and are never returned to the browser. Reads and writes are scoped to the accounts you connect, every mutation is recorded in the audit log, and you can revoke any integration or key at any time. See the Privacy page for details.
Whatever you write in. Plans, reports and the agent's own memory all match your language — describe a goal in English, 中文, 日本語 or others and get the same plan and the same guardrails.
No installation. Sign up, create your organization, connect an account on the Integrations page via OAuth, then write a goal in plain language in the Workspace. The orchestrator plans the steps and waits for your approval on anything that writes. The full guides live in the Documentation.
Multi-agent automation across SEO, analytics, ads and publisher platforms — Search Console, GA4, Google Ads, DV360, CM360, Ad Manager, AdSense, AdMob and BigQuery, plus 250+ more apps via Composio and your own MCP. Bring an account, write a goal in plain language, and approve what ships.