How the tools work
The parameters every tool shares, what a response carries, pagination and async tasks.
Manifold's tools are sorted into groups by where the data comes from. Every tool looks something up, so your agent can call one again whenever it needs fresh data, and a monitor repeats the calls on a schedule and reports what changed. Your agent picks the tools for you, so you rarely need to know them by name, but this page explains what they all have in common. Each group has its own page with every tool's parameters, result and cost.
Common parameters
Many tools take some of these parameters. A tool's page lists the ones it takes.
| Parameter | Default | What it does |
|---|---|---|
location | "United States" | Where the search runs, by name or provider code |
language | "en" | ISO 639-1 language code |
device | "desktop" | "desktop" or "mobile" |
limit | 100 | The most rows to return, up to 1,000 |
cursor | None | The next page: meta.cursor from the previous result |
provider | The server's choice | Leads tools only: which data provider to use. meta.provider says which one answered |
max_credits | None | Paid tools: refuse the call with BudgetExceeded if the estimate is higher |
dry_run | false | Paid tools: return only meta with the estimate. Costs nothing |
Search data is from Google only.
Targets
Tools that take a target accept a domain or a URL, and the form you give sets how much of the site it covers:
| Target | Covers |
|---|---|
example.com | The domain and all its subdomains |
www.example.com | That host only |
https://example.com/pricing | That page only |
What a response carries
A result is the tool's data plus two extras: meta, which describes the call, and hints, which suggests what to do next. The same JSON also comes back as text, for clients that read text only.
meta field | What it says |
|---|---|
provider | The data provider that answered |
credits_charged | What this call cost |
est_credits | The estimate made before the call |
cached | true if the answer came from the cache |
data_as_of | When the provider's data is from |
rows_returned, rows_available | Rows in this result, and in total where the provider says |
cursor | Pass it back as cursor for the next page. null on the last page |
request_id | Quote it when you contact support |
hints lists up to three tools worth calling next, cheapest first, each with its estimate.
In the data, null means the provider has no data, and 0 means it measured zero. So a keyword with a null volume is one the provider has no number for, which is not the same as one nobody searches for.
Pagination
List tools return up to limit rows and a meta.cursor. To get the next page, call the same tool with the same input and cursor set to that value. The cursor is opaque, so pass it back exactly as you got it.
Batch inputs
Some tools take a list, such as a list of domains. They return the rows that worked and an errors list for the rest, and each error names its input. One bad domain does not fail the whole batch.
Async tasks
Two tools take longer than a request should: seo_run_technical_crawl and aeo_run_ai_answers. Instead of making your agent wait, they return a task_id and poll_after_s at once. Your agent can wait that many seconds, or call get_task with the task_id straight away:
{"task_id": "task_01J..."}While the task runs, get_task holds the call for up to 25 seconds, then returns TaskPending, and your agent calls it again at once. When it is done, it returns the result under result, with the settled charge in meta.credits_charged. get_task is free, and results are kept for 30 days.
Feedback
When a tool returns an error that does not explain itself, or a result that looks wrong, your agent can report it with submit_feedback. It takes a kind (error, wrong_data, feature_request or other), a message, and optionally the tool and the request_id from the call's meta, which lets us find that exact call:
{"kind": "error", "tool": "seo_get_serp", "request_id": "3f1c9a2e-8d4b-4c6f-9e1a-7b2d5c8f0a13", "message": "seo_get_serp returned ProviderUnavailable five times in ten minutes."}It returns a feedback_id. The report reaches only the Manifold team, by email, and nothing comes back through the tool. You can also ask your agent to send us anything else, such as a tool you wish existed. submit_feedback is free, and each workspace can send 20 reports a day. Like get_task, it has no group on the Tools page, so it is always there.
Once per workspace, after it has been used on a few different days, one result carries feedback_request in its meta. Your agent finishes what you asked, then asks you once whether you have feedback on Manifold, and sends your answer with prompted: true only if you give one. ChatGPT and Codex get no request.
Tool capabilities
Each tool is labelled Read, Write or Sensitive write for connector review. The labels match the MCP annotations (readOnlyHint and destructiveHint) the server advertises.
- Write: creating or updating a monitor, submitting a sitemap or URLs to Google or Bing, and
submit_feedback. - Sensitive write: deleting a monitor.
- Read: every other tool, including crawls and AI-answer runs that only look data up.
Skills
The tools answer one question each. For a whole job, such as an SEO plan or a lead list, install Manifold's marketing skills. Each skill covers one area of marketing: what works there now, with dated sources, and which tools to call when the job needs numbers. That page shows how to install them in Claude Code, Codex, Gemini CLI, Cursor and other agents. Start with manifold-get-started: it asks about your goal and picks the skills to run.