Loading ToolBench
A declarative DSL and compiler for reproducible gene-editing design, exposed as an MCP server over stdio with JSON-RPC 2.0.
Helix defines 4 tools with complete input schemas and descriptions. Tool names are action-oriented (helix_reference, helix_compile, helix_resolve_gene, helix_list_models). Descriptions are substantive (100-250 chars) and explain WHAT, WHEN, and WHY. All parameters have type definitions and descriptions. However, output schemas are NOT documented, LLMs cannot see what fields to expect from compile results, gene resolution, or model listings. Error handling is mentioned in helix_compile description ('clean error with line/col') but no structured error recovery guidance. Parameter descriptions are good but lack examples of valid enum values inline (e.g., 'ensembl' vs 'ncbi' are listed but not explained). No tool annotations (readOnlyHint, destructiveHint, idempotentHint) despite all tools being read-only operations.
Compile a Helix spec into a ranked, safety-screened gene-editing design. This is the core tool: the deterministic compiler turns your spec into guides, delivery, primers, safety verdict, feasibility, and provenance. Returns a clean error (with line/col for syntax) if the spec is invalid.
List the simulation/computation backends registered for each role (sequence, on-/off-target, repair, base-/prime-edit, variants) — the selectable `simulate {}` options. External ones (Cas-OFFinder, inDelphi, BE-Hive, PRIDICT) activate only when their binary/package is installed on the backend.
Return the Helix DSL grammar, an authoring guide, and worked example specs. Call this FIRST so you know how to write a valid spec before compiling.
Resolve a gene symbol to an edit-window DNA sequence and genomic coordinates via a live database (Ensembl/NCBI). Needs network access. Use the returned sequence inline in a spec, or the locus for variant-aware design.
Output schemas not documented. LLMs cannot see what fields helix_compile returns (e.g., top_guides, feasible, safety.allowed, infeasibility_notes are mentioned in authoring guide but not in tool schema). This forces agents to guess at response structure and breaks downstream tool chaining.
No tool annotations. All 4 tools are read-only (no state mutation), but readOnlyHint is not set in tool definitions. This prevents agents from optimizing retry logic and understanding safety boundaries.
Parameter descriptions lack inline enum explanations. helix_resolve_gene accepts database enum [ensembl, ncbi, fallback] but does not explain when to use each (e.g., 'ensembl for human genes, ncbi for cross-species'). Agents must infer intent.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 62 | 2026-07-28+ | v2 |
Error recovery guidance incomplete. helix_compile mentions 'clean error with line/col' but does not guide LLM on next steps (e.g., 'If syntax error, review GRAMMAR.md and retry'). Agents lack actionable recovery paths.
helix_resolve_gene network dependency not surfaced in error handling. Tool description says 'Needs network access' but does not document timeout behavior, fallback strategy, or what to do if Ensembl/NCBI is unreachable. Agents cannot plan for failure.