Open Source · MIT License

Typed interoperability for
AI agent tools

AI capabilities are in the pre-TypeScript era. Effector adds a typed interface layer — 40 standard capability types, static composition checking, and cross-runtime compilation from a single sidecar manifest.

Get Started → Try the Playground

$ npm init @effectorhq/effector-code

GitHub

Capability Types

Compile Targets

Tests Passing

Dependencies

The Problem

Every AI tool has a type system. It's just invisible.

Your tool takes something, produces something, and needs credentials from the environment. Today these interfaces are implicit. Effector makes them explicit.

Without types

Chain two tools, pray they work

Search tools by keyword

Read the README to understand a tool

Manually test every combination

Rewrite per runtime

No permission model

With effector

Type-check composition before execution

Discover capabilities by interface type

Machine-readable interface contracts

Automated compatibility verification

One manifest, every runtime

Declared + audited security

How It Works

One file. Four capabilities.

Drop an effector.toml next to your tool. Your implementation stays untouched — the manifest is pure metadata.

Declare your interface

Pick from 40 standard types to describe what your tool accepts, produces, and requires. Declare permissions explicitly. One file, no code changes.

effector.toml

[effector]
name = "code-review"
version = "1.0.0"

[effector.interface]
input = "CodeDiff"
output = "ReviewReport"
context = ["Repository"]

[effector.permissions]
network = false
filesystem = ["read"]

Validate statically

Catch type errors, missing permissions, and structural issues at build time. The validator checks your manifest against the 40-type catalog and JSON schema.

terminal

$ npx @effectorhq/core validate .
  ✓ effector.toml parsed
  ✓ Schema valid
  ✓ Types: CodeDiff → ReviewReport

$ npx @effectorhq/audit check .
  ✓ Permissions match source
  ✓ No undeclared access

Compile to any runtime

One manifest, every target. The compiler reads your effector.toml and emits MCP schemas, OpenAI function definitions, LangChain configs, or JSON IR.

compile.js

import { compile } from '@effectorhq/core/compile';

// Same manifest, different targets
const mcp  = compile(def, 'mcp');
const oai  = compile(def, 'openai-agents');
const lang = compile(def, 'langchain');
const ir   = compile(def, 'json');

Compose with confidence

Chain tools into pipelines with type-level guarantees. The composition checker proves output types match input types — incompatible wiring is caught before any API call.

pipeline

// Static type checking for tool pipelines

git-fetch        → CodeSnippet
    ↓
code-review     → ReviewReport    ✓
    ↓
slack-notify    → SlackMessage    ✓

// Score: 1.0 — all types compatible

Benchmark

Verified against established benchmarks.

Two-tier evaluation grounded in BFCL, τ-bench, ToolSword, MCPToolBench++, and API-Bank. Not custom metrics — dimensions that published research has shown correlate with tool-use accuracy.

Type Resolution

87.7%

Manifest Validation

89.7%

Composition Safety

100%

Security Detection

81.4%

Compilation

100%

Toolchain Score / 100

179

Test Cases

Known Gaps

3.2ms

Runtime

Dependencies

$ node benchmark.js --verbose

Known Limitations (why not 100%)

Type Checker Wildcard matching uses naive prefix — BarCode matches Code* incorrectly

Validator Does not reject unknown top-level fields or check interface types against catalog

Security Regex-based — misses leetspeak evasion, WebSocket exfiltration, produces false positives on code comments

Baseline MCP effector MCP Δ

Comparable Both formats can express these

Function Selection Signal

BFCL · Berkeley Function Calling Leaderboard

+28

Parameter Extraction Signal

BFCL + API-Bank · Li et al., ACL 2023

−14

Schema Completeness

MCPToolBench++ · Nexus (Srinivasan et al., 2023)

+37

Differential Capabilities effector adds to MCP

Multi-Step Composition Safety

τ-bench (Yao et al., 2024) + MCP-Bench

+75

Safety & Permission Coverage

ToolSword (Ye et al., 2024) + SafeToolBench (Guo et al., 2024)

+53

+18

Comparable Δ

+64

Differential Δ

+36

Combined Δ

8/8

Composition Chains

⚠

D2 Regression: compile() generates inputSchema.properties only for envRead variables. 6/10 tools show parameter constraint regression. This is a real compiler limitation, not a measurement artifact.

Composition Chain Verification

✓ code-review → slack-notify exact-match

✓ security-scan → slack-notify subtype

✗ file-search → deploy-service incompatible

✗ db-query → summarize-doc incompatible

Baseline tools cannot participate — MCP has no typed interface model

Grounded in BFCL, MCPToolBench++, MCP-AgentBench, τ-bench (Yao et al., 2024), ToolSword (Ye et al., 2024), SafeToolBench (Guo et al., 2024), API-Bank (Li et al., ACL 2023).
Deterministic · No API keys · Reproducible: node benchmark.js --verbose

Adoption

Zero friction. Zero dependencies.

Sidecar manifest

Your tool's implementation stays untouched. effector.toml sits next to your code like .d.ts sits next to .js.

5 minutes to start

npx create-effector scaffolds everything. One command to validate, one to compile. No configuration.

Zero dependencies

The entire toolchain uses only Node.js built-ins. No node_modules tree. Installs in under a second.

Build-time only

No runtime overhead. Effector runs at build and CI time. Zero code in the request path.

Ecosystem

9 focused packages. One shared kernel.

Every package builds on @effectorhq/core. Every package uses only Node.js built-ins.

Core

@effectorhq/core

Shared kernel — TOML parser, type checker, schema validator, compiler, CLI

11 modules · 69 tests

Core

effector-types

40-type catalog — fields, aliases, subtypes, frequency scores

40 types · 40 tests

Dist

Everything runs in the browser.

No sign-up. No download. No setup.

Start with one file.
Ship to every runtime.

One command to scaffold. One manifest to declare your interface. One compiler for every target.

Get Started → Read the Docs

$ npm init @effectorhq/effector-code

Typed interoperability for
AI agent tools

Every AI tool has a type system. It's just invisible.

One file. Four capabilities.

Declare your interface

Validate statically

Compile to any runtime

Compose with confidence

Verified against established benchmarks.

Zero friction. Zero dependencies.

Sidecar manifest

5 minutes to start

Zero dependencies

Build-time only

9 focused packages. One shared kernel.

@effectorhq/core

effector-types

create-effector

@effectorhq/skill-lint

@effectorhq/audit

@effectorhq/compose

@effectorhq/skill-eval

@effectorhq/graph

openclaw-mcp

Everything runs in the browser.

Playground

Studio

Graph Explorer

Type Catalog

Start with one file.
Ship to every runtime.

Typed interoperability for AI agent tools

Every AI tool has a type system. It's just invisible.

One file. Four capabilities.

Declare your interface

Validate statically

Compile to any runtime

Compose with confidence

Verified against established benchmarks.

Zero friction. Zero dependencies.

Sidecar manifest

5 minutes to start

Zero dependencies

Build-time only

9 focused packages. One shared kernel.

@effectorhq/core

effector-types

create-effector

@effectorhq/skill-lint

@effectorhq/audit

@effectorhq/compose

@effectorhq/skill-eval

@effectorhq/graph

openclaw-mcp

Everything runs in the browser.

Playground

Studio

Graph Explorer

Type Catalog

Start with one file.Ship to every runtime.

Typed interoperability for
AI agent tools

Start with one file.
Ship to every runtime.