Fold

Fold is an Effect-native, provider-agnostic, isomorphic agent core with an optional opinionated coding agent, CLI, and TUI. It supports multiple model providers (including openai/anthropic-compatible and subscription-backed providers for Codex, Grok, and OpenCode Zen), subagents, and RLM-like orchestration patterns.

npm install -g @humanlayer/fold
foldcode                         # open the interactive TUI
foldcode --prompt "fix the tests" # noninteractive; framed assistant output goes to stdout; useful for CI
foldcode auth --help             # configure and manage authentication

foldcode auth ... handles provider authentication. The separately published @humanlayer/fold-cli package is an optional Node/headless CLI; it does not provide the TUI.

• fold: skeleton package for CLI distribution
• fold-core: the isomorphic core with log-based state, tool support, hook support, subagents and skill facades, event streaming, auto-compaction, session management
• fold-agent: opinionated agent with built-in profiles, filesystem tools (read, write + edit for claude models / apply_patch for codex ones, bash, skill, agent, web_search and web_fetch); sane hook configuration
• fold-cli: CLI & TUI package
• fold-codex: effect codex provider
• fold-opencode: effect opencode zen provider
• fold-tui-theme: theme tokens and example app
• fold-xai: effect XAI provider

Isomorphic agents with fold-core

The ergonomic API is descriptor-based: define an agent, then run it against a session backend. No Effect Layer, Toolkit, or runtime wiring is exposed.

import { anthropicModel, defineAgent, startSession } from '@humanlayer/fold-core'
import { Config, Effect } from 'effect'

export const ask = (prompt: string) =>
	Effect.gen(function* () {
		const apiKey = yield* Config.redacted('ANTHROPIC_API_KEY')
		const agent = defineAgent({
			name: 'assistant',
			model: anthropicModel({ model: 'claude-sonnet-4-6', apiKey }),
			systemPrompt: 'Be concise and helpful.',
		})
		const session = yield* startSession({ agent }) // in-memory by default
		return yield* session.send(prompt)
	}).pipe(Effect.scoped)

defineTool, model descriptors, and event-log descriptors use the same data-first API, so hosts can supply browser, worker, or filesystem implementations without changing the agent definition.

Custom coding agents with fold-agent

Compose fold-agent's platform tools and JSONL backend with the same fold-core API. codingTools({ cwd }) installs the complete tool union; fold-core selects the advertised tools from the active model on every request: Claude gets write/edit, while GPT/Codex gets apply_patch. Switching models reselects the tools automatically.

import { codingTools, jsonlEventLog } from '@humanlayer/fold-agent'
import { anthropicModel, defineAgent, startSession } from '@humanlayer/fold-core'
import { Config, Effect } from 'effect'

const program = Effect.gen(function* () {
	const apiKey = yield* Config.redacted('ANTHROPIC_API_KEY')
	const cwd = '.'
	const agent = defineAgent({
		name: 'reviewer',
		model: anthropicModel({ model: 'claude-sonnet-4-6', apiKey }),
		systemPrompt: 'Review the project. Make changes only when explicitly asked.',
		tools: codingTools({ cwd }),
		autoCompact: { enabled: true },
	})
	const session = yield* startSession({ agent, cwd, log: jsonlEventLog('.fold/review.jsonl') })
	return yield* session.send('Find the highest-risk code in this project.')
}).pipe(Effect.scoped)