Documentation

All 19 Tools at a Glance

*configure_optimizer: Enterprise settings (policy_mode, audit_log, config lock, session_retention_days) require Enterprise tier.

The main entry point. Analyzes a raw prompt, detects ambiguities, compiles an optimized version, scores quality, and estimates cost across providers. Returns a PreviewPack for review.

PreviewPack — Contains session_id, quality_before (0–100), compiled_prompt, blocking_questions, assumptions, cost_estimate, model_recommendation, compilation_checklist, and changes_made.

Quick pass/fail check of a prompt. Returns a quality score, the top issues detected, and a suggestion for improvement. No compilation, no session. Ideal for CI gates or lightweight checks.

Quality score (0–100), pass/fail verdict based on configured strictness threshold, top 2 issues, and an improvement suggestion.

Classifies a prompt by task type (13 types), reasoning complexity (6 levels), dimensional risk score, and suggests an optimization profile. Useful for understanding a prompt before committing to optimization.

taskType, complexity (simple_factual / analytical / multi_step / creative / long_context / agent_orchestration), riskScore (0–100), riskDimensions (underspec, hallucination, scope, constraint), suggestedProfile, and signals.

Routes to the optimal model based on task complexity, risk, budget, and latency preferences. Returns a full recommendation with a decision_path audit trail showing every routing decision. Two-step routing: (1) complexity + risk determine the default tier, (2) budget/latency overrides produce the final tier.

ModelRecommendation — Contains primary (model, provider, temperature, maxTokens), fallback, confidence (0–100), costEstimate, rationale, tradeoffs, savings_vs_default, and decision_path array.

Full pre-flight analysis in a single call: classifies task, assesses risk, routes model, and scores quality. Returns a complete decision bundle. Counts as 1 metered use (does NOT call optimize_prompt internally).

Task classification, complexity result, risk score with dimensions, model recommendation with decision_path, quality score, cost estimate, and pre-flight deltas (estimated token savings from compression and pruning).

Refines a prompt by answering blocking questions or providing manual edits. Re-runs the full analysis pipeline and returns an updated PreviewPack.

Updated PreviewPack with re-scored quality, re-compiled prompt, and reduced blocking questions.

Sign-off gate. Approves the compiled prompt and returns the final optimized prompt ready for use. Fails if blocking questions remain. In Enterprise enforce mode, runs a policy enforcement check before allowing approval.

The final compiled prompt, quality score before compilation, and the compilation checklist summary.

Estimates token count and cost across all supported providers for any prompt text. No session needed. Covers 11 models from Anthropic, OpenAI, Google, and Perplexity.

CostEstimate — Contains input_tokens, estimated_output_tokens, per-model costs array (model, provider, input/output cost in USD), recommended_model, and recommendation_reason.

Pricing covers models across Anthropic, OpenAI, Google, and Perplexity. Updated regularly. See Models for supported models.

Compresses context (code, docs) by removing irrelevant sections while respecting protected zones (fenced code, tables, lists, JSON, YAML). Multi-stage pipeline intelligently removes boilerplate while preserving meaningful content.

compressed_context, removed_sections, original_tokens, compressed_tokens, tokens_saved, savings_percent, and mode.

Scores and ranks MCP tools by relevance to a task intent. In prune mode, marks the lowest-scoring tools for removal to save context tokens. Respects mention protection and the always-relevant tool set (search, read, write, edit, bash).

Ranked tool list with relevance_score (0–100), signals, and tokens_saved_estimate per tool. Plus pruned_count, pruned_tools, and total tokens_saved_estimate.

Configures optimizer behavior. Supports mode, threshold, strictness, default target, ephemeral mode, session limits, and Enterprise features (policy mode, audit log, config lock with passphrase protection).

The full OptimizerConfig object and list of applied_changes.

Returns current usage count, plan limits, remaining quota, tier information, and first/last used timestamps.

total_optimizations, limits (TierLimits), remaining (lifetime + monthly), tier, first_used_at, last_used_at.

Returns aggregated optimization statistics: total optimized, total approved, average quality score, top task types, blocking question frequency, and estimated cost savings in USD.

total_optimized, total_approved, average_score_before, task_type_counts, blocking_question_counts, estimated_cost_savings_usd.

Activates a Pro, Power, or Enterprise license key. Uses Ed25519 asymmetric signature verification (public key only, offline, zero dependencies). The license payload contains tier, issued_at, expires_at, and license_id — no PII.

status ("activated"), tier, expires_at, license_id, and limits for the activated tier.

Checks current license status, tier, and expiry. Re-validates expiry on every read. Returns purchase URLs if no license is active.

LicenseData (or null with purchase links if inactive): tier, issued_at, expires_at, license_id, valid, validation_error (if expired).

Lists all optimization sessions with metadata (no raw prompts). Returns newest-first, with task type, quality score, state, and prompt hash.

SessionListResponse — Array of SessionRecord (session_id, state, task_type, quality_before, prompt_hash, target) plus total_sessions and storage_path.

Exports full session details including raw prompt, compiled prompt, and reproducibility metadata. Auto-calculates rule_set_hash, rule_set_version, risk_score, and custom_rules_applied.

SessionExport — Contains raw_prompt, compiled_prompt, quality_before/after, rule_set_hash, rule_set_version, and metadata (target, task_type, complexity, risk_score, custom_rules_applied, engine_version, policy_mode, policy_hash).

Deletes a single optimization session by ID. Audit-logged in Enterprise tier.

deleted: true and session_id, or a not_found error.

Bulk-purges optimization sessions by age policy or deletes all. Safe-by-default: requires explicit parameters. Supports dry_run preview, keep_last protection, and config-based session_retention_days fallback.

PurgeResult — deleted_count, retained_count, scanned_count, deleted_session_ids (capped at 100), dry_run, cutoff_date, effective_older_than_days.