Actoviq Agent SDK
DeconBear/actoviq-agent-sdk
SDK TypeScript para agentes multi-tool, multi-sessão com integração MCP e benchmark
Ver no GitHubSDK experimental em TypeScript para construir agentes multi-ferramenta e multi-sessão com API limpa, integração MCP e helpers de memória. Suporta 4 modos de cooperação entre modelos, workflows dinâmicos orquestrados via JavaScript e busca web com TavilySearch.
Instalação
npx convez-ai DeconBear/actoviq-agent-sdk1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Construir agentes inteligentes que usam múltiplas ferramentas em sequência
- Orquestrar workflows complexos com cooperação entre modelos de IA
- Integrar agentes com Model Context Protocol (MCP) servers
Veredito

SDK robusto e bem-documentado para construir agentes sophisticados em TypeScript. Destacam-se os modos de cooperação entre modelos e suporte a MCP, embora ainda em desenvolvimento ativo (v0.5.0) com APIs que podem evoluir.
README
A tradução para o português deste README ainda está na fila.
Actoviq Agent SDK
Documentation site: https://deconbear.github.io/actoviq-agent-sdk/
Actoviq Agent SDK is an experimental TypeScript SDK for building multi-tool, multi-session agents with a clean-first public API, MCP integration, and memory helpers.
This project is inspired by excellent agent projects and runtimes including Claude Code, Codex, Deepagents, and similar work in the ecosystem. Actoviq remains an independent project with its own public SDK surface and documentation.
v0.5.0 Highlights
- Model Team — 4 multi-model cooperation modes (Panel/Router/Discussion/Executor-Reviewer), per-member independent provider config, $ENV_VAR apiKey resolution, global AgentPool
- Dynamic Workflows — JS script-based multi-agent orchestration with
agent()/parallel()/pipeline()primitives, sandboxed runtime, schema enforcement - Worktree Tools —
EnterWorktree/ExitWorktreewith stack-based cwd,.worktreeinclude, PR checkout, hooks for non-git VCS - TavilySearch — AI-optimized web search, pure TypeScript, auto key detection
- Standard Benchmark — Self-contained framework with DeepSeek judge, HTML dashboard, 4-agent comparison (Hadamard/Bridge/Official)
- TUI/REPL:
/workflows,/worktree,/teamslash commands
This repository is still under active development. APIs and runtime behavior may continue to evolve. Issues and pull requests are very welcome.
Install
npm install actoviq-agent-sdk zod
For local examples, place your config at:
~/.actoviq/settings.json
You can also preload a custom JSON file with loadJsonConfigFile(...).
Quick Start
import { createAgentSdk, loadDefaultActoviqSettings } from 'actoviq-agent-sdk';
await loadDefaultActoviqSettings();
// Default: Anthropic protocol
const sdk = await createAgentSdk();
// Or use OpenAI / OpenAI-compatible APIs (DeepSeek, vLLM, etc.)
const sdk = await createAgentSdk({
provider: 'openai',
baseURL: 'https://api.deepseek.com',
model: 'deepseek-chat',
});
try {
const result = await sdk.run('Introduce yourself in one short sentence.');
console.log(result.text);
} finally {
await sdk.close();
}
Run the repository examples with:
npm run example:actoviq-quickstart
npm run example:actoviq-agent-helpers
CLI REPL
After installing the package, you can start an interactive scrollback-mode REPL directly from the terminal:
npx actoviq-react [work-dir]
This launches a readline-based interactive agent with:
- Real-time streaming output in the main terminal buffer (native scrollback)
- Tab completion for slash commands, including session model, permission, compact, and resume controls
- Command history via ↑↓ arrow keys
- Ctrl+C to abort the current request, press twice to exit
Note: actoviq-react is a lightweight scrollback REPL, not a full-featured TUI. It does not use an alternate screen buffer, ScrollBox, or rich terminal rendering. It is designed for quick interaction and debugging. For the full terminal UI, use actoviq-tui below.
Terminal UI (TUI)
actoviq-tui is the full terminal UI for the Hadamard SDK, modeled on Claude Code's REPL design: the transcript prints into native scrollback while a redrawable bottom region hosts the status line, a Claude-style prompt bar, the slash-command menu, and permission dialogs.
npx actoviq-tui [work-dir] [options]
# Options
# --config <path> Load a specific Actoviq settings JSON file
# --permission-mode <mode> default | acceptEdits | plan | bypassPermissions (default)
# --model <model> Override the configured model
# --resume <session-id> Resume a stored Hadamard SDK session
# --continue Continue the most recently updated session
Features:
- Streaming transcript in native scrollback — assistant text,
⏺ Tool(args)calls, and⎿ ✓/✗result lines flush into the normal terminal buffer; scrollback and copy/paste work as usual. - Live status line — spinner, elapsed time, tool count, context-size estimate, and the current tool while the agent works.
- Claude-style prompt bar — type
\thenEnter(orCtrl+J) for a newline;↑/↓walk input history; the caret renders inline. - Slash-command menu — type
/to open a filtered menu (↑↓select,Tabcomplete,Enterrun)./resumeopens a searchable project-session picker, while/resume <session-id>still resumes directly. - Runtime catalogs —
/skills,/agents,/mcp, and/pluginsbrowse the Hadamard SDK capabilities visible to the current workspace;/helpprovides a searchable command reference. - Model and effort controls —
/modelopens a model picker,/model configedits the provider, masked API key, base URL, andmin/medium/maxmodel tiers, and/effortselectslow,medium,high,max, or automatic provider behavior. - Dream controls —
/dreamopens a run/status picker;/dream runand/dream statusremain available for direct use. - Mid-run steering — keep typing while the agent works and press
Enter: the message is queued and injected into the very next model request (shown as⧗ queued). - Permission dialogs — with
--permission-mode default, mutating tools pause for an approve / always-allow / deny dialog. Always-allow choices are stored with the session and restored on resume. - Interrupts —
Escaborts the current run;Ctrl+Cclears the input (twice quickly exits);Ctrl+Dexits on an empty prompt. - Context management built in — the Hadamard SDK auto-compacts long sessions mid-run and reactively recovers when a provider rejects an oversized prompt; compactions surface as
∿ context compactednotices.
Both CLIs share the same Hadamard SDK runtime defaults (Actoviq settings from ~/.actoviq/settings.json, core tools, bypassPermissions, uncapped tool iterations) and run against any Anthropic-compatible or OpenAI-compatible provider.
By default, Hadamard SDK sessions are scoped to the current workspace under ~/.actoviq/projects/<workspace-key>. Explicit sessionDirectory settings still take precedence.
Model tiers are provider-neutral aliases. Configure them with ACTOVIQ_DEFAULT_MIN_MODEL, ACTOVIQ_DEFAULT_MEDIUM_MODEL, and ACTOVIQ_DEFAULT_MAX_MODEL, then use min, medium, or max anywhere a model can be selected.
Tutorials
- English tutorial: docs/en/README.md
- Chinese tutorial: docs/zh/README.md
- GitHub Pages docs site:
Start with these examples:
Contributing
Contributions are welcome. If you spot a bug or a documentation gap, please open an issue or submit a pull request.
Licensed under the MIT License.