RENKIN

Plan routes. Audit routes. Keep the evidence local.

RENKIN is a pure-Rust computer-aided synthesis planning (CASP) engine and a tool-neutral route-audit layer. It searches from target SMILES to a declared stock, then can audit routes from RENKIN, AiZynthFinder, Syntheseus, and SynPlanner with the same deterministic checks.

Current release: v1.0.14.

Start here

BrowserPythonCLI

Open the local WebAssembly playground

pip install renkin

import renkin
routes = renkin.find_routes("CC(=O)Oc1ccccc1C(=O)O", depth=5)

cargo install renkin
renkin --target "CC(=O)Oc1ccccc1C(=O)O" --depth 5

Two complementary jobs

Job	RENKIN does	Read next
Plan	Searches with bounded A*/beam exploration, rule/template expansion, and declared building-block stock	Quick start
Audit	Checks route topology, structure, stock policy, optional forward replay, and evidence receipts	Audit reproducibility

The core has no C/C++ dependency and is available as a Rust crate, CLI, Python package, MCP server, and browser WebAssembly module. Local audit inputs and private-stock policies stay local unless an operator deliberately exports them.

The default planner currently has 24 hand-crafted rules. Its repository stock file contains 402 compounds; installations without that file, and WASM, use a compiled-in 152-compound fallback. Supply an explicit stock file when this distinction matters to a run.

What an audit report means

An audit verdict is pass, fail, or partial. partial is intentional: if the supplied route cannot support a check, for example because it lacks an atom-mapped reaction representation, RENKIN records not_evaluable instead of inventing a result. A route report is not an experimental-success, yield, safety, or regulatory claim.

Useful guides:

• Private stock policy
• Chemical review rubric
• Evidence-carrying interchange
• MCP server

Benchmark boundary

The registered Phase 55 TEST used a frozen 690-target cohort, shared stock, and declared budgets. RENKIN had 481/690 strict routes (69.71%) and AiZynthFinder 4.4.1 had 32/690 (4.64%); the paired coverage difference was +65.07 percentage points (95% CI +61.45 to +68.55).

This supports a coverage result only for that registered configuration. It does not establish universal planner superiority or experimental viability. Peak-RSS and time-to-first-route receipts are still pending, so it does not support a whole-cohort performance claim. See the Benchmark overview and formal Phase 55 result record.

Choose an interface

Need	Interface
Application integration	Python API or Rust API
Browser-only workflow	WASM / JavaScript
Automation agent	MCP server
Forward validation and retrieval	Forward tools

For source, releases, and issue tracking, visit github.com/kent-tokyo/renkin.