A Windows automation tool whose if: expressions survive a paste from Slack

The short answer is who the user is. Power Automate Desktop, UiPath, Blue Prism, and AutoHotkey are aimed at business users, RPA developers, or power users writing scripts by hand. Terminator is a developer framework. You install it as an npm package, a pip package, a Rust crate, or an MCP server that Claude Code, Cursor, Windsurf, or VS Code can drive. The product ships with a Playwright-shaped selector API, a workflow engine with typed variables, a recorder that emits YAML, and a tiny expression language for conditional steps. The expression language, documented below, is one of the least-discussed differentiators and the reason human-authored workflows survive copy-paste.

crates/terminator-mcp-agent/src/expression_eval.rs, 458 lines. The public entry is fn evaluate(expression: &str, variables: &Value) -> bool at line 34. It calls normalize_expression first, then evaluate_internal on the cleaned string. The parser handles &&, ||, !, ==, !=, >, <, >=, <=, and four built-in functions: contains, startsWith, endsWith, and coalesce. There is also an always() function that takes no arguments and returns true, used by GitHub-Actions-style step gates.

It rewrites eight Unicode characters into their ASCII equivalents before the parser ever runs. Curly single quotes U+2018 and U+2019 become straight '. Curly double quotes U+201C and U+201D become straight ". Backticks become single quotes. Non-breaking space U+00A0, thin space U+2009, and narrow no-break space U+202F all become regular ASCII space. Then the expression is trimmed. The function is seven lines of chained .replace() calls at the top of the file. You can read it in under 30 seconds.

People copy expressions out of Slack threads, Notion docs, Google Docs, Word, and ChatGPT conversations. Every one of those sources will silently replace your straight quotes with curly quotes, especially on macOS where the default input substitution does it without asking. An expression like if: quote_type == 'Face Amount' will look identical on screen but the character codes are different. Terminator's interpreter normalizes the input so the workflow runs. Most other automation tools' conditional syntaxes fail with a cryptic parse error and leave the user to hunt for an invisible character.

In YAML, attach an if: field to any step. Example: - tool_name: click_element; arguments: { selector: 'role:Button && name:Save' }; if: "contains(env.product_types, 'FEX') && env.retry_count < 3". The expression is evaluated against the current workflow variables before the step runs. If it returns false, the step is skipped and the next step in the sequence is tried. The if: field is inspected by helpers.rs at lines 281 and 340, which call expression_eval::evaluate(inner_str, variables). Real workflow fixtures in the repo use exactly this pattern, including if: contains(product_types, 'FEX') at helpers.rs line 1214.

Four data functions plus one sentinel. contains(array_or_string, 'value') tests for membership or substring. startsWith(string_var, 'prefix') and endsWith(string_var, 'suffix') are self-explanatory. coalesce(x, y, z, default) returns the first truthy value from its arguments, or the last argument as a literal if nothing is truthy. always() with no arguments returns true and is used as a step gate. Logical operators && || ! compose these. Binary comparison ==, !=, >, <, >=, <= works on any JSON value with smart type coercion between strings, numbers, and booleans.

Terminator handles undefined variables explicitly at lib lines 183 through 191. For equality operators, undefined is never equal to anything, so undefined == 'x' returns false and undefined != 'x' returns true. For numeric comparisons, undefined is treated as less than any value, so undefined > anything is false and undefined < anything is true. This makes defensive expressions like if: env.retry_count < 3 work on the first run when retry_count does not exist yet. Other platforms often throw, which means you have to wrap every step in a try/catch or pre-seed every variable.

Yes. The whole project is MIT licensed on GitHub at mediar-ai/terminator. The expression_eval.rs file is ~450 lines, has no external dependencies beyond serde_json and tracing, and ships with a full test module starting at line 351 covering coalesce with nulls, missing variables, empty strings, zero-as-falsy, and more. If you want to fork the parser for your own tool, the file is self-contained and easy to extract.

Power Automate Desktop is a drag-and-drop designer built on the Robin expression language, aimed at business analysts. Conditions are authored in a property grid. Terminator is a code-first framework: you write a YAML workflow (or drive it from TypeScript, Python, Rust, or an MCP tool call) and your LLM coding assistant can read and edit it the same way it edits source files. The if: field is the same whether you are building interactively, recording a session, or having Claude generate the workflow. The expression interpreter is optimized for paste-friendly developer input, not for a designer surface.

If you already use Cursor, Claude Code, VS Code, or Windsurf, the fastest path is the one-line MCP install: claude mcp add terminator "npx -y terminator-mcp-agent@latest". Your coding assistant can then open apps, fill fields, click buttons, and run multi-step workflows. If you want to build automations programmatically, the NodeJS and Python SDKs are both on npm and pypi under the terminator name. If you want to embed the engine in a Rust binary, the crate is terminator-rs on crates.io. Everything reads the same if: expression format.