AGENT Mode
AGENT is Genie’s main workspace. ASK, CODE, QUERY, and DATA are capabilities inside AGENT rather than separate modes. Describe the outcome you want; AGENT interprets the request, binds only relevant FileMaker context, and chooses the safe tool and response path.
AGENT is included in Free Core. Coding Access adds native script authoring, executable DATA plans, and other reviewed coding actions.
What AGENT can do
Section titled “What AGENT can do”- Explain, inspect, debug, and review scripts, calculations, fields, tables, relationships, layouts, and FileMaker behavior.
- Read an exact current, named, or attached script when the request makes it relevant.
- Use read-only Query for summaries, tables, and charts.
- Run one exact existing script after a fresh native confirmation.
- Give complete human-readable script steps and calculation fragments.
- With Coding Access, prepare native script Preview/Copy/Create/Apply artifacts.
- With Coding Access, prepare a sealed DATA Review & Execute card.
- Use provider-native web search where the configured provider supports it.
The local tool set
Section titled “The local tool set”Embedded AGENT and external Local MCP share eight public tool contracts:
fm_session— current FileMaker session, slim context manifest, metadata refresh, and safe file actionsfm_search— progressive discovery of live names, guides, schemas, and referencesfm_read— one exact script, attachment, field definition, metadata target, or guidefm_query— read-onlySELECTdatafm_evaluate— non-mutating FileMaker calculation evaluation and diagnosticsfm_stage— validate, compile-preview, or stage reviewed script, DATA, and layout artifactsfm_workbench— inspect or manage staged work and coordinated ChangeSetsfm_run— status or confirmed start of one exact existing script
External Local MCP requires Coding Access. Embedded AGENT keeps its Free read-only and confirmed-script-run capabilities without exposing the external server.
Context is pull-only and relevance-bound
Section titled “Context is pull-only and relevance-bound”AGENT receives a slim manifest first: active script identity, focused or selected one-based lines, attachment names and handles, and a context revision. Script, document, and image bodies are loaded only through an exact read when needed.
- An open script remains available but is not automatically the target of every turn.
- Referring to “this script,” the current line, an exact recent script name, or a trusted selection can bind the target.
- Ambiguous target identity fails closed and may produce one clarification.
- Stale attachment revisions are rejected instead of silently reading changed content.
- Images become provider-native image input only after an explicit attachment read.
Questions and investigation
Section titled “Questions and investigation”Question-only work stays read-only. AGENT may inspect live evidence before it answers, but it does not turn an explanation into a hidden change. Continue in the same conversation when you want to move from understanding to implementation.
Script work
Section titled “Script work”For a new script, an open or attached script can be sealed as reference evidence while the new target remains empty. For an edit, AGENT must resolve one exact edit target.
- Free Core returns complete numbered FileMaker steps, options, calculations, and manual implementation guidance.
- Coding Access can validate a structured script artifact and present the clean native review card.
- Genuine compile or structural faults produce a warned draft.
- Missing numeric reference IDs do not block a draft when FileMaker can resolve the target by name.
- Preview, Copy, Create, and Apply remain explicit user actions.

Read-only Query
Section titled “Read-only Query”Ask for a table, chart, summary, or raw result in the prompt. AGENT resolves exact schema identity before it uses read-only SQL. Query is included in Free Core and cannot mutate records or schema.

Reviewed DATA changes
Section titled “Reviewed DATA changes”DATA writes are a Coding Access capability inside AGENT. AGENT must compile one exact native write plan; only the sealed validation result becomes a Review & Execute card. Composer text cannot confirm execution, and external MCP cannot Execute. The card action plus a fresh native confirmation are both required.
Script Workspace Agent
Section titled “Script Workspace Agent”The compact Script Workspace Agent is the same AGENT runtime in a smaller surface. It remains bound to the exact Script Workspace that opened it even while the composer owns focus. Opening it does not silently copy the script or refresh selection.
Use Refresh Selection only when you intentionally want to recapture the Script Workspace selection. QuickCode remains a separate bounded Coding Access shortcut for its explicit sealed selection.
Provider route and clarifications
Section titled “Provider route and clarifications”Settings exposes one Coding provider/model route. AGENT uses that route for its sealed tool loop. Legacy Logistics preferences are compatibility mirrors only.
AGENT asks at most one authority question per task. Buttons carry stable decisions such as new script, edit script, or read-only; typed answers are interpreted against the sealed question and task context.