Markifact - Ads & Analytics

Search for available Markifact operations that can be executed by the authenticated user. Workflow: find_operations → get_operation_inputs → run_operation (read-only) or run_write_operation (when requires_approval: true). After selecting an operation_id, always call get_operation_inputs next. Never call run_operation / run_write_operation without first inspecting the schema via get_operation_inputs. Input shapes, nested objects, and field names vary per operation and cannot be guessed. Returned operation IDs are registry-backed Markifact operation IDs, not arbitrary action names. Returns JSON `{results, count}`. Each result has operation_id, app_id, app_name, description and requires_approval (true means it changes data and must run through run_write_operation).

find_operations

Get a signed URL for a private file. Use this when an operation returns a `file_id` and you need a temporary, directly accessible URL. Operation `file_url` values are user-facing app URLs that require Markifact authentication; call this tool to generate a signed URL that the model or external tools can open directly. The signed URL expires after 1 hour. Returns the signed URL as plain text, ready to use or share with the user.

get_file_url

Get the input schema, required fields, defaults and constraints for an operation. Required step between find_operations and run_operation/run_write_operation. Call this once per operation_id before executing it — schemas contain nested objects, account selectors, and field names that vary per operation and cannot be guessed. If you already retrieved the schema for this operation_id earlier in the conversation, reuse it instead of calling again. Only pass fields from this schema in run_operation/run_write_operation input_data for the same operation_id. Returns the operation's input JSON Schema, or `null` when it takes no inputs.

get_operation_inputs

List saved Markifact connections for an app/platform. A connection is a saved authenticated integration in Markifact. The returned `id` is the `connection_id` used to choose which saved login/integration an operation should use. Important: - `connection_id` is not an external platform account ID. - External account IDs, customer IDs, property IDs, and similar values are separate operation inputs. - One connection may have access to multiple external accounts. - In most cases, you don not need to provide `connection_id`. If the user has only one connected login for a platform in their workspace, operations usually use that connection automatically. Use this tool only when the user wants a specific saved integration or has multiple connections for the same platform. Returns JSON with available connections containing `id`, `type`, and `display_name`.

list_connections

Create a signed upload URL for a media file selected in the upload UI. Returns upload details including file_id, file_url, upload_url and upload_token, or a validation error.

prepare_media_upload

Inspect a file returned by a previous operation as numbered lines of text. Use this when an operation returns a `file_id` and you want to inspect the file contents before deciding what to do next. This is mainly useful for exported JSON, CSV, logs, and similar files produced by Markifact operations. It returns the file as numbered text lines so you can quickly review the structure, schema, headers, sample rows, or specific sections in chunks. Returns up to `limit` numbered lines and a footer showing how many lines remain. Use `offset` to read the next chunk. After inspecting, you can use {{file-id}} in other tool inputs to pass the full parsed content for processing.

read_file

Submit actionable product feedback to Markifact with the user's authorization. When you find a likely Markifact bug, missing capability, or product limitation, tell the user and offer to report it. You may draft the report for their review. Call this tool only after either: 1. the user explicitly asks you to submit the feedback, or 2. you offered to report a discovered issue and the user explicitly agreed. If the request already authorizes submission, do not ask again. Never use this tool to ask permission. Submissions are irreversible and include the feedback, operation IDs, and the account's contact and reference identifiers. After success, briefly tell the user it was submitted. Report only issues Markifact itself caused. Low-level operations such as gads_mutate or meta_ads_update_adset pass requests straight to the platform, so their responses are the platform's, not Markifact's. A write the platform accepted is complete; if the state differs later, something else changed it. If you are unsure how a platform behaves, research its documentation. Never report third-party policies, outages, rate limits, rejected content, campaign performance, or user configuration. Before reporting a missing capability, use find_operations to check for a dedicated or generic operation that supports it. Never run a write just to test whether a capability exists. Include only the minimum non-sensitive evidence and summarize errors instead of copying raw data. Never include passwords, API keys, tokens or other auth details, payment information, raw conversation history, full request or response payloads, or unrelated personal or customer data. Avoid duplicate reports. Use `other` only when the feedback does not fit another category. Write the title and description in English. Do not promise that Markifact will respond or implement a fix. Returns JSON confirming whether the submission succeeded.

report_markifact_feedback

Execute a **read-only** Markifact operation, such as fetching reports, listing records, selecting accounts, or looking up account details. Use this in direct response to the user's current request. Do not use it for startup checks or background context gathering. For write/create/update/delete actions, use `run_write_operation` instead. Operations returned by `find_operations` with `requires_approval: true` MUST be executed via `run_write_operation`. operation_id must be an exact Markifact operation ID returned by find_operations for the authenticated user. Never call this without first inspecting the schema via `get_operation_inputs`. Input shapes, nested objects, and field names vary per operation and cannot be guessed. input_data must contain only fields from the get_operation_inputs schema for the same operation_id. Reporting tips - Reporting operations (GA4, Google Ads, Meta Ads, etc.) require account IDs. Use the relevant `select_accounts` operation first, treating user-provided account names as substring matches. - Never guess metric or dimension values. Use the relevant `list_report_fields` operation to confirm available fields before building a report. - For Google Ads, always prefer using the `gads_get_report` operation. Use `gads_run_gaql_query` only for advanced or unsupported reporting queries. Files - Large results return a `file_id` plus a small `data_preview`. - Reference a file in other operation inputs as `{{file-id}}`, e.g. `{"data": "{{file-xe3fw}}"}`, exactly as returned; never modify or invent file ids. - Markifact loads the file automatically: JSON becomes a dict/list, CSV/XLSX a list of row dicts, images/binary raw bytes. No parsing needed. - The file ID is for you to move data between operations efficiently. When sharing results with the user, always share links using the `get_file_url` tool. Returns JSON with the operation result.

run_operation

Execute a **write** operation that creates, updates, or deletes external data. Use this tool ONLY for operations marked with `requires_approval: true` in `find_operations` results. These are destructive actions like creating campaigns, updating ad statuses, sending emails, etc. operation_id must be an exact Markifact operation ID returned by find_operations for the authenticated user. **APPROVAL REQUIRED**: Every call writes irreversibly to the user's live external accounts. Before each invocation, summarize the exact change in plain language (operation, target account/resource, key field values) and wait for explicit user confirmation. Never call without user's confirmation, approval is CRITICAL. Never call this without first inspecting the schema via `get_operation_inputs`. Input shapes, nested objects, and field names vary per operation and cannot be guessed. input_data must contain only fields from the get_operation_inputs schema for the same operation_id. Write-operation tips - Always prefer a dedicated operation when one already matches the task. - Some platforms have a generic fallback operation for create, update, or remove mutations not covered by a dedicated operation, for example `gads_mutate` for Google Ads. Use the fallback before reporting a missing write feature. Files - Large results return a `file_id` plus a small `data_preview`. - Reference a file in other operation inputs as `{{file-id}}`, e.g. `{"data": "{{file-xe3fw}}"}`, exactly as returned; never modify or invent file ids. - Markifact loads the file automatically: JSON becomes a dict/list, CSV/XLSX a list of row dicts, images/binary raw bytes. No parsing needed. - The file ID is for you to move data between operations efficiently. When sharing results with the user, always share links using the `get_file_url` tool. Returns JSON with the operation result.

run_write_operation

Upload an image or video asset for later operations. Use this when the user needs to provide media that will be reused in later tool calls, for example creating Facebook ad image or video asset, or a Google Ads creative flow. The tool returns file URLs that can be passed into later operations that accept uploaded media. Only image and video uploads are supported.

upload_media