Skip to content

Choosing an agent platform

Soulacy overlaps with agent frameworks, visual automation products, and personal assistant gateways, but these products solve different layers of the stack. This page compares their operating models rather than declaring a universal winner.

Research date: September 10, 2026. Claims below are linked to first-party documentation or repositories. Product capabilities and packaging change, so follow the links before making a purchasing or architecture decision.

The one hard reason

Nothing risky happens unless you approved it, or you can point to the line that allowed it.

Every other option makes what an agent may do a matter of trust: code someone wrote, a canvas someone configured, or a service in someone else's cloud. In Soulacy it is a property of the runtime. Agents fail closed by default, and every exception is a line in a YAML file you can read. So for any action an agent ever takes, you can answer one of two questions: who approved it, or which line permitted it.

Do not take our word for it. Check it in five minutes on your own machine:

  1. Ask for something risky. Write a small agent and ask it to delete a file. It cannot: system tools are not even offered until the agent's ID is listed in runtime.allow_system_agents in your config and the agent declares capabilities: [system]. It is a per-agent list, not a yes/no switch: stricter than a boolean, on purpose.
  2. Turn them on and ask again. Now the agent stops and waits for your approval before the privileged step runs.
  3. Schedule it for 3 a.m. The same step is refused because nobody is there to approve it, until the agent's file says unattended: true.
  4. Read the diff. Every change that made the agent more capable is a line you wrote in a file you can read, review, and roll back. Safe Undo covers the changes it makes.

The simple version

Soulacy is a private system that runs your AI agents. You describe an agent in one file. It runs on a computer you control. You reach it from your phone or your chat apps. A privileged action needs a recorded approval or an explicit agent and gateway policy grant.

Everywhere else Soulacy
Where it runs Code frameworks leave you to build and host the application; visual builders need a server stack; hosted agent services keep your agents in their cloud One binary on a laptop, a $5 VPS, or a Raspberry Pi, with the web UI, login, channels, and schedules already inside it
How you describe an agent A program you write, or a canvas you drag boxes around A single YAML file anyone can read, review in a pull request, and roll back; Studio can write it from plain English and repair it when it breaks
What keeps it safe Approvals and guardrails you wire in per project Enforced by the runtime for every agent: confirmations for risky tools, the intent gate against injected instructions, a sandbox for tool code, and Safe Undo for reversible changes

Building agents into a product of your own? A code framework or vendor SDK fits better. Want a drag-and-drop canvas for business workflows? A visual platform fits better. Soulacy is for people who want their own agents running privately, reachable from their phone, without building or renting the system around them. The rest of this page covers each product and the operating details behind that summary.

At a glance

Product Primary shape Authoring model Deployment and operations Human oversight and security Mobile and channel surface
Soulacy Self-hosted agent runtime and control plane Versionable SOUL.yaml plus Studio One Go binary or container with GUI, runs, schedules, logs, diagnostics, and delivery Confirmation approvals, capability policy, prompt-injection scanning, intent gate, audit, and production-readiness checks Native iPhone companion plus HTTP, webhook, Telegram, Slack, Discord, WhatsApp, email, Teams, and Google Chat
LangGraph Low-level orchestration framework/runtime Python or JavaScript graphs and functions Run in your application; LangSmith adds managed, hybrid, or self-hosted deployment and operations Durable checkpoints and configurable interrupts support approve, edit, reject, and custom review flows Client and channel experiences are composed by the application team
CrewAI Multi-agent framework with Crews and Flows Python; Crew Studio adds a visual/low-code surface in CrewAI AMP Self-managed Python application or CrewAI AMP deployment and observability Guardrails, event/state handling, and human-feedback patterns are framework/platform features Integrations and end-user clients are selected by the application team
OpenAI Agents SDK Code-first agent SDK Agents, tools, handoffs, guardrails, sessions, and structured outputs in application code Ships inside your application; built-in tracing records model, tool, handoff, and guardrail events Input, output, and tool guardrails can block or replace calls; the application supplies its review UX Client and channel experiences are application-defined
Claude Agent SDK / Managed Agents Code SDK plus a managed agent execution service Code for Agent SDK; versioned model, prompt, tool, MCP, and skill configuration for Managed Agents Agent SDK is process-operated; Managed Agents supports Anthropic-managed or self-hosted workers Permission policies support allow/ask decisions and tool-confirmation events; custom-tool enforcement remains the application's responsibility Client and channel experiences are application-defined
Dify Visual LLM application platform Visual agents, Chatflows, Workflows, knowledge, and plugins Dify Cloud or a multi-service self-hosted Docker Compose stack; includes logs and monitoring surfaces Workflow Human Input nodes can pause for review; moderation and provider controls are configurable Published web apps, APIs, and embedding are first-class surfaces
n8n Visual workflow automation platform with AI nodes Node canvas, expressions, code nodes, and AI Agent nodes n8n Cloud or self-hosting through npm, Docker, Compose, and cloud guides; execution history and debugging are built in Selected AI tools can require human approval through a configured review channel Delivery is assembled from workflow integrations and front ends
Flowise Visual platform for agents and LLM workflows Assistant, Chatflow, and Agentflow builders Self-hosted or deployed across common cloud/container platforms; tracing, analytics, and evaluations are built in Agentflow checkpoints support durable human input and per-tool approval APIs, SDKs, and an embedded chatbot are documented product surfaces
OpenClaw Self-hosted personal-assistant gateway CLI/onboarding, gateway configuration, skills, plugins, and Control UI A long-running Node.js gateway/daemon owns sessions, auth, channels, and state Pairing, allowlists, tool policy, sandboxing, exec approvals, and openclaw security audit; its documented trust model is one trusted boundary per gateway Broad chat-channel catalog plus official iOS and Android companion nodes

What the categories mean

Frameworks: LangGraph, CrewAI, and the model-vendor SDKs

Frameworks give software teams composable primitives and leave substantial product assembly to the adopter. LangGraph explicitly positions itself as a low-level runtime for long-running, stateful agents and documents durable execution, streaming, memory, and human-in-the-loop. Its persistence layer supports checkpoints, fault recovery, replay/time travel, and review; its interrupt model can pause and resume a run for approve, edit, or reject decisions. LangSmith deployment adds managed, hybrid, and self-hosted operational choices.

CrewAI centers on Crews and Flows in Python. Its documentation describes guardrails, memory, knowledge, and observability as framework capabilities, while CrewAI AMP adds managed deployment, monitoring, and Crew Studio. This makes CrewAI a closer comparison for code-first multi-agent applications than for a packaged private gateway.

The OpenAI Agents SDK provides agents, tools, handoffs, sessions, and structured outputs. It includes input, output, and tool guardrails and built-in tracing, but the surrounding application, deployment, end-user UI, and channel adapters remain product decisions for the adopter.

Anthropic now documents two related shapes. The Claude Agent SDK is operated as part of an application, while Managed Agents can use Anthropic-managed or self-hosted workers. Managed Agent versions capture the model, prompt, tools, MCP servers, and skills, and permission policies can always allow or ask before supported tool use. Anthropic explicitly notes that applications remain responsible for enforcing permissions around their custom tools.

Visual platforms: Dify, n8n, and Flowise

Dify is a broad visual platform rather than only a workflow library. Its open-source repository lists visual workflow building, agents, RAG, model management, observability, APIs, and plugins. It offers cloud hosting and a Docker Compose self-hosting path. Current Dify documentation also includes workflow history and logs and a Human Input node.

n8n is primarily a general workflow-automation product with AI integrated into its node canvas. Official documentation describes Cloud and self-hosted options, npm, Docker, Compose, and cloud-provider installation paths, an AI Agent node, human approval for selected AI tools, and execution debugging.

Flowise focuses on visual generative-AI construction. Its official overview lists Assistant, Chatflow, and Agentflow builders alongside tracing, analytics, evaluations, human-in-the-loop, APIs, SDKs, and an embedded chatbot. Agentflow V2 documents checkpointed human input and tool approval, and Flowise provides several deployment targets.

Personal-assistant gateway: OpenClaw

OpenClaw is the closest product-shape comparison in this list. It runs a self-hosted gateway that owns sessions, authentication, channels, and state; its getting-started guide installs a Node.js gateway/daemon and opens a browser Control UI.

It has a substantial first-party channel and device story:

  • The channel catalog includes Telegram, Slack, Discord, WhatsApp, Google Chat, iMessage, SMS, WebChat, and more.
  • The iOS companion connects to a gateway, exposes opt-in device capabilities, caches recent conversations, and supports durable offline message queuing.
  • The Android companion supports gateway pairing and review of pending command approvals.

OpenClaw also documents security controls rather than leaving them implicit: pairing and allowlists, tool policies, and sandboxing, exec approvals, and a structured security audit. Its documentation is equally explicit about the boundary: OpenClaw assumes one trusted operator or team per gateway and does not claim hostile multi-tenant isolation.

Where Soulacy is different

Soulacy packages a self-hosted runtime, versionable agent definition, browser control plane, and native iPhone operations surface together. The goal is to operate multiple private agents without first building the deployment console, run history, schedules, approvals, delivery adapters, and production diagnostics around a framework.

The differentiators documented in this project are:

  • A reviewable SOUL.yaml remains the canonical agent artifact, with Studio providing a guided authoring surface.
  • The security stack combines capability scope, confirmation approvals, untrusted-content handling, prompt-injection signals, an intent gate, audit records, SSRF controls, and launch-readiness checks.
  • The channel runtime includes HTTP, webhooks, Telegram, Slack, Discord, WhatsApp, email, Teams, and Google Chat, with mapping and delivery diagnostics in the same control plane.
  • Soulacy for iPhone is an operator and delivery client: it can connect to a remote gateway, chat, review approvals, watch runs, inspect activity, and receive mobile-channel output.
  • The distribution can run as a single binary or container on a laptop, private host, or VPS; operators still own TLS, host security, backups, credential rotation, and network policy.

OpenClaw is stronger when the desired product is a personal assistant spanning a large catalog of chat networks and device-node capabilities. Visual platforms are stronger when a canvas and connector ecosystem are the center of the job. Frameworks are stronger when agent behavior is embedded deeply in a custom software product. Soulacy is aimed at teams that want a private, multi-agent operations layer with versionable configuration and conservative production checks.

How to evaluate any option

Run one representative task end to end before adopting a platform:

  1. Use the actual model provider, tool, trigger, and delivery destination.
  2. Exercise both least-privileged and high-impact actions.
  3. Test a human-review pause and confirm the run resumes after restart.
  4. Test bad credentials, unavailable providers, timeouts, duplicate delivery, and an unreachable channel.
  5. Confirm what is retained in logs, traces, chat history, and audit records.
  6. Verify backup, restore, upgrade, and rollback on the intended host.
  7. Test the real phone or chat client over the same network path users will use.

For Soulacy, start with the Quick Start, then use Studio to build that representative agent.

Primary sources