Loading ToolBench
MCP server for neurochemically-grounded narrative engineering
Advanced Writer MCP exhibits significant quality gaps across tool definitions. While 18 tools are registered with names and descriptions present, the descriptions are inconsistent in depth and clarity. Schemas are partially visible in the provided code but many tool definitions appear to be inferred rather than directly visible in the source (e.g., tools defined in app/api/run/route.ts or other referenced files not provided). Parameter descriptions vary widely in specificity. Most tools lack structured output schema documentation. Error handling guidance is absent. The server implements HTTP transport (good) and exposes prompts + resources (good), but lacks modern spec patterns like tool annotations, structured error recovery, or per-request metadata handling. Conservative evaluation applied given inference of some tool definitions and absence of complete source visibility.
Builds the next draft version from the StoryScope review. Non-destructive and auto-incrementing (v1->v2->v3...); re-running when the latest version has no review of its own REBUILDS that version from its predecessor (rerun/resume — the predecessor is never touched). The planner assigns each critique issue to EXACTLY ONE operation: 'rewrite', 'line_edit' (surgical anchored edits), 'global_line_edit' (manuscript-wide line pass), 'cut_scene', 'merge_scenes', or 'add_scene'. Every change is (a) checked against the World Bible's hard rules (unresolved contradictions are surfaced loudly and ledgered), (b) VERIFIED against its own directive (PASS/FAIL with cited evidence, retried with auditor feedback), (c) guarded against unrequested prose loss (a rewrite that sheds more words than its directive authorizes is REVERTED), (d) scored for neurochemical/pathology deltas (before->after), and (e) logged with deterministic diff stats. Ends with a COVERAGE REPORT mapping every critique item -> op -> scene -> verified status (including items it could NOT action, honestly), and updates the persistent cross-version issue ledger. Executes the saved plan from storyscope-reports/<version>/revision-plan.json (compiled by storyscope_final_review; user-editable in the Studio) when present; auto-plans otherwise; 'directives' overrides both. This tool only revises PROSE — for canon reconciliation use reconcile_storyscope_canon.
Systematically revises identified narrative pathologies (character inconsistencies, plot holes, pacing issues, neurochemical misalignments) across multiple scenes in batch mode.
Generates a batch of varied story concepts and premises with configurable wildness, seed topics, and exclusion filters.
Missing output/response schemas for all tools. Tool definitions lack documented return types, field names, and structures. LLMs cannot predict what data they will receive or plan downstream tool chains.
No error handling guidance. Tool descriptions do not explain what errors can occur, whether they are retryable, or what the LLM should do next. Errors will be silent failures or unhandled exceptions.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 66 | 2026-07-28+ | v2 |
Constructs a comprehensive world bible documenting setting, rules, history, magic systems, character relationships, and hard constraints that all narrative revisions must respect.
Check the status/result of a background job started by running a long tool with async=true. Returns running | completed | failed plus the final summary.
Compiles a detailed, executable revision plan from StoryScope critique (executive summary and specialist lens reports), assigning each issue to a specific operation (rewrite, line_edit, cut_scene, etc.), cross-checking against the world bible, and prioritizing by severity. The plan is saved as JSON and can be edited by the user before apply_storyscope_revisions executes it.
Continues an existing narrative by generating new scenes that maintain character consistency, emotional tracking, and story momentum using affect logs and neurochemical scoring.
Creates a complete narrative from a story premise, including scenes, character arcs, and dialogue. Builds the draft manuscript structure with neurochemically-grounded emotional beats.
Develops detailed character profiles with neurochemical archetypes, emotional range, motivations, and relationship dynamics.
Expands a short story, outline, or synopsis into a full-length novel manuscript with complete scenes, dialogue, and neurochemically-grounded emotional arcs.
Deterministic find and replace across manuscript text with preview mode and safe application tracking.
List recent background jobs (most recent first) with their status.
Publishes or exports the final manuscript in multiple formats (PDF, EPUB, HTML) with formatting, metadata, and optional distribution integration.
Reconciles world-bible contradictions and canon inconsistencies flagged during StoryScope review, updating the bible and optionally patching affected scenes.
Performs comprehensive narrative review scoring pacing, character consistency, emotional arcs, and story structure coherence.
Rewrites an individual scene with specified narrative objectives, preserving character voice and emotional continuity.
Selects and applies narrative structure templates (Hero's Journey, Save the Cat, Three-Act Structure, etc.) to existing story concept.
Performs comprehensive StoryScope review using specialist lens reports (actors table, style analysis, pacing, neurochemical tracking) to generate critique and revision directives. Returns executive summary and detailed lens reports for each aspect.
Many tools defined outside of visible source code (app/api/run/route.ts, referenced files not provided). Cannot verify actual schema, parameter types, or implementation details. Tool definitions treated as inferred, capping individual scores at 50.
Async operations (create_narrative, continue_narrative, batch_revise_pathologies, expand_to_novel, storyscope_final_review, apply_storyscope_revisions) support background execution with job IDs but lack documented timeout behavior, retry semantics, or status field descriptions.
Parameter descriptions lack specificity on constraints. For example, 'wildness' in brainstorm_ideas lists no enum values; 'format' in publish_story has no explicit list of supported formats; 'directive' in rewrite_scene and apply_storyscope_revisions lack format guidance.
Tool names like 'batch_revise_pathologies' are domain-specific jargon ('pathologies') that may not be self-documenting for LLMs unfamiliar with narrative engineering terminology. Consider clearer naming like 'batch_fix_narrative_issues'.
No tool annotations (readOnlyHint, destructiveHint, idempotentHint) present. LLMs cannot determine which tools are safe to retry, which modify state, or which require confirmation before execution.
apply_storyscope_revisions has exceptionally complex nested parameters (directives array with enum op values, merge_with/after_scene conditionals) but parameter relationships are documented only in descriptions, not enforced via schema constraints. LLMs may pass invalid combinations.
Pagination support missing from tools that return lists. review_narrative, compile_revision_plan, storyscope_final_review, check_job, list_jobs lack limit/offset parameters and do not document result counts.
No confirmation/dry-run support for destructive operations. publish_story (export), batch_revise_pathologies, apply_storyscope_revisions (rewrite multiple scenes), and reconcile_storyscope_canon (modify world bible) lack confirmation patterns to prevent accidental irreversible changes.