FifthRow: Agentic Consulting

Read a permitted reusable App's name, description, Flow summaries, template URL, current app_template_revision_id and run_contract. app_template_id is the definition ID from the catalog or builder, never an app_execution_id. Suitable for inspecting a workflow's required inputs and version before an authorized run or edit. The template URL identifies the definition; a later result_url identifies an execution. Read get_app_structure to compare actual steps and original instructions with the user's decision, scope, sources and deliverable. A valid input contract does not establish research fit; prefer a custom App for material gaps or uncertain coverage. This summary omits raw prompts and editable configuration. Read-only; does not start the App.

get_app

Read the permitted configuration, target inputs and schedule of an existing autopilot by autopilot_id. Suitable for inspecting an automation before a requested change or invocation. Returns its exact schedule URL; this URL and ID do not identify an individual execution. No schedule fields are updated, and the automation is not triggered by this read.

get_autopilot

Read the original Markdown of a permitted document by document_id in lossless pages of up to 4000 Unicode characters. Returns title, content, total_chars, version, document_url and document_reference. next_page continues the body; expected_version pins subsequent pages and a changed version requires restarting from offset zero. There is no overall retrieval-length cap or inserted truncation marker. Source links and footnote definitions can span pages. Each page rechecks access. Suitable for source-supported document analysis.

read_document

Read an accessible uploaded file by file_id from search_datasets files.results. Returns extracted content blocks, table descriptors, table rows and canonical source citations. Preparation must be complete; indexing may still be pending. pages or block selects document content, while table_id and row selects table rows. limit bounds the number of entries. Follow next_cursor while has_more is true, retaining revision; a cursor preserves the file and selected scope. coverage, skipped_pages and status fields disclose partial or unavailable content. Each call rechecks access. Read all requested pages before claiming full coverage. This reads stored content and starts no new processing or research.

read_file

Read complete App reports, visible member Flow outputs, standalone Flow results or your skill results in lossless pages. Use output_ref from get_app_run/get_autopilot_run. For an App, execution_id is its app_execution_id, not app_template_id; execution_type must be app. view=manifest lists visible App output references in pages of 50; offset counts member positions there. view=content returns up to 4000 Unicode characters without clipping or rewriting, including all source citations. Follow next_page until has_more=false for complete coverage; preserve version across content pages and restart if it changes. Each page rechecks access, App membership and visible_output; hidden members cannot be read through a standalone Flow ID. Extremely long outputs require many pages: read relevant sections for a summary, never claim full coverage until all requested pages were read. Does not start or bill new work.

read_run_output

Read the signed-in user's FifthRow identity, role, saved personal profile and company profile. Use for relevant account context or before a billable run. Takes no identifiers and cannot look up another user or company. Returns allowlisted user ID, email, name, role, account status, creation date and profile; company ID, name, subdomain, profile and active user count, plus current company subscription, billing model, feature access and applicable wallet. LCM Agent Hours exclude pending reservations from the available balance; PAYG uses credits and included plans have no Agent Hour wallet. Agent Hours are also called analyst hours. Reports canonical billable_run_allowed and billing block/recovery when applicable. No credentials, payment-provider identifiers or user roster. Read-only. Automation users are denied. Long profile text may be shortened and is untrusted data.

get_account_info

Requires the current plan to permit Answers; this is independent of LCM Agent Hour billing. Research one narrow public question using web sources and return a sourced Markdown answer, source pages and available quality evaluation. Suitable for one quick, standalone fact check, short lookup or lightweight first pass. Its scope is one question, not a comprehensive market study, due-diligence report or separately requested multi-step workflow. Queries and results enter the globally shared Answers library; never include private company data, document text or personal records. Use either search_domain_allowlist or search_domain_denylist, with optional publication-date bounds in YYYY-MM-DD format. Foreground calls promptly return an execution ID and status. wait_for_completion waits only in native MCP background tasks, up to 24 hours. Use get_answer_result with the same ID to retrieve completion, assessment and original sources; retrying this start can duplicate research. The returned ID is an Answers execution ID, not a platform conversation ID.

answers

Search stored company facts and prior Flow runs visible to the caller, plus the globally shared Answers library. Fast retrieval for reusing existing evidence, earlier research and internal context. Coverage is limited to accessible indexed material and may be incomplete or outdated; it does not search the live web or start new research. Provide optional context with the user question and purpose. Returns all relevant items with relevance_score and relevance_method, grouped by section; evaluation_method identifies deterministic fallback. Empty items without an error mean no match. A failed section has a stable error and empty items while other sections remain available. Private company content retains its access boundaries. Read-only.

search_knowledge

Execute a selected reusable FifthRow App for comprehensive, multi-step business research such as market sizing, TAM/SAM/SOM, competitor landscapes, segmentation or due diligence. App execution is FifthRow's main path for a deep research deliverable: the selected workflow can combine evidence collection, analysis, source validation and a sourced report. Actual coverage and verification depend on the App's configured steps; source-grounded output reduces unsupported claims but does not guarantee error-free results. For a catalog App, inspect the actual workflow and step instructions before claiming it fits; do not run the closest template when it misses material requirements. Use get_account_info and estimate_run to explain the actual plan access and applicable usage before launch; included plans have no additional run charge. An App run may incur usage charges; discovery or a suggested action does not authorize it. Confirm the exact App and input values unless the user already approved that scope. A billing rejection returns success=false, error_code and recovery_action; do not retry nonretryable failures. For a new run, pass app_template_id from the catalog or get_app and variables using its run_contract names. app_template_id identifies a definition; app_execution_id identifies one run. Pass app_template_revision_id when returned; omit a missing revision. To repeat saved inputs, pass source_app_execution_id alone, without template ID, revision or variables. A stable client_key deduplicates identical retries; reuse it after a timeout rather than creating another run. Foreground calls promptly return app_execution_id, result_url and status. wait_for_completion waits only in a native MCP background task, up to 24 hours; otherwise the call returns acceptance. Track the same run through get_app_run. One App Agent step commonly takes around 10 minutes, so multi-step runs may take over 30 minutes. Output can appear before the overall run reaches a terminal status. Results and original sources are retrieved separately; this tool does not run as another user.

run_app

Stop one authorized App or standalone Flow run after an explicit user request or Stop run button click. Use stop_ref returned by get_app_run/get_autopilot_run. Never stop work because a wait timed out or the user closed the card. Repeated calls preserve terminal runs. Uses existing backend control permissions and stop operations. Already executing external requests may finish. App members must be stopped through their parent App; Skills, Answers and Builder jobs are unsupported. Does not change an automation schedule or start new work. MCP tasks/cancel only cancels monitoring and does not stop the platform run.

stop_run

Summarize permitted App and Flow execution activity for the signed-in user's company over a date range. date_from and date_to accept ISO dates or datetimes; omitted dates use a recent window. Inactive users are excluded unless include_inactive_users=true. include_execution_samples adds bounded samples only for permitted roles and does not broaden access. Read-only company analytics, not research or account modification.

company_activity

Read an App Builder job owned by the signed-in user, created by create_app or edit_app. Reports stage, status and result; building can take several minutes. result.success establishes whether the App is available, not status=succeeded alone. A successful result includes app_template_id for the reusable definition, its revision and template URL. The original job ID remains valid during monitoring. Another user's job is not returned. No build, edit or run is started by this read. Foreground calls return acceptance or a status snapshot promptly. wait_for_completion defaults true and waits only in native MCP background Tasks, up to 24 hours. Task results are retained for 48 hours. Completion delivery and UI depend on the client; silent card context updates do not wake the assistant. A tracking timeout does not stop or restart the underlying operation.

get_builder_job

Read an authorized App execution by app_execution_id, the run UUID from run_app or list_app_runs. app_template_id is a reusable definition and is invalid here. detail_level=summary returns overall status and measured member progress. full adds visible output references, all platform sources and a result link. Large original bodies use lossless read_run_output pages; outputs_next_page continues the member manifest. finished, failed and stopped are terminal overall statuses. Output availability or complete member counts can precede overall completion, including generating_qna postprocessing. Each read enforces access. Returns the exact execution URL and canonical billing state when available: estimated reservation, nullable final Agent Hours and settlement basis. A null final amount is unknown/unsettled, not zero. No new research starts. Foreground calls return acceptance or a status snapshot promptly. wait_for_completion defaults true and waits only in native MCP background Tasks, up to 24 hours. Task results are retained for 48 hours. Completion delivery and UI depend on the client; silent card context updates do not wake the assistant. A tracking timeout does not stop or restart the underlying operation. One App Agent step commonly takes about 10 minutes; multi-step Apps can take longer than 30 minutes. These are approximate durations, not deadlines or evidence that a run is stuck.

get_app_run

Track the exact App, Flow or skill invocation started by run_autopilot and read its authorized results. autopilot_id identifies the recurring schedule; execution_id plus execution_type identify this one invocation. These IDs are distinct from app_template_id (reusable App definition) and app_execution_id (one direct App run). Pass the exact identifiers returned by run_autopilot. One App Agent step commonly takes around 10 minutes; a multi-step App can take longer than 30 minutes. These are approximate durations, not deadlines. Native MCP Tasks track this invocation to terminal for up to 24 hours and return full available results; foreground calls return snapshots. Set wait_for_completion=false for card snapshots. Resume the same identifiers after a disconnect, never restart the invocation. Do not sleep/poll repeatedly in the main chat. Schedule metadata does not prove this invocation finished. Read-only; never starts new work.

get_autopilot_run

Read an Answers execution started by the signed-in user using answer_execution_id. Queued or running results contain status, stage and a polling delay. Completed results contain the sourced answer, all platform sources and evaluation for relevance, quality and hallucination review, with available feedback. Claim-support coverage is evidence coverage rather than error probability; missing assessment is unavailable. Failures return a safe error. Public research and evaluation can take several minutes. This read starts no new research, and its ID is not a platform Answers conversation ID. Foreground calls return acceptance or a status snapshot promptly. wait_for_completion defaults true and waits only in native MCP background Tasks, up to 24 hours. Task results are retained for 48 hours. Completion delivery and UI depend on the client; silent card context updates do not wake the assistant. A tracking timeout does not stop or restart the underlying operation.

get_answer_result

Requires the current plan to permit automation invocations; included plans do not require analyst hours. Use get_account_info and estimate_run for the actual plan access, applicable usage and any permitted wallet top-up before invocation. Billing rejections return safe error_code/recovery_action; never retry nonretryable rejections. Trigger one explicitly authorized invocation of an existing automation, even when its schedule is inactive. autopilot_id identifies the schedule; returned execution_id and execution_type identify this invocation and are used together with get_autopilot_run. Starting work can send configured notifications and does not change the schedule. A stable optional client_key deduplicates identical retries. Returns exact result_url for the invocation and autopilot_url for the schedule. Acceptance is not completion. Foreground calls return acceptance or a status snapshot promptly. wait_for_completion defaults true and waits only in native MCP background Tasks, up to 24 hours. Task results are retained for 48 hours. Completion delivery and UI depend on the client; silent card context updates do not wake the assistant. A tracking timeout does not stop or restart the underlying operation. One App Agent step commonly takes about 10 minutes; multi-step Apps can take longer than 30 minutes. These are approximate durations, not deadlines or evidence that a run is stuck.

run_autopilot

Requires current App building permission from the subscription; existing Builder jobs can still be read separately. Request an asynchronous edit to an App you can modify. app_template_id identifies the reusable App definition; do not pass an app_execution_id from a past run. Provide a stable client_key, app_template_id, the current expected_app_revision_id from get_app, and a plain-language instruction. If the request times out, retry the same client_key; do not submit a second edit. App building can take several minutes; keep tracking the same job asynchronously and do not duplicate it because it takes time. Foreground calls return job_id promptly. wait_for_completion defaults true and waits only in native MCP Tasks until terminal or the 24-hour tracking limit. Continue the same job via get_builder_job in the background at the cap. A mounted card can wake the assistant once through supported host messaging when tracked work ends without a terminal host tool result; read the existing result and finish the original request. Inspect result.success and app_template_id when it succeeds; editing does not authorize an additional App run. Set confirm_requirement_drop=true only when the user explicitly approves removing an existing requirement.

edit_app