Feature matrix
Every platform in one table. Facts are drawn from each vendor’s own materials; any cell we couldn’t confirm from a primary source is flagged (unverified).
18 platforms. Cells marked (unverified)aren’t confirmed from a primary source.
| Platform | Score | Pricing | Hosting model | Memory & state | MCP | Integrations | Model providers | Open source |
|---|---|---|---|---|---|---|---|---|
| AlfeManaged | 3.3 | Freemium — the site describes a tenant-wide credit pool that funds compute, model usage, voice, channels, and storage, plus optional per-agent subscriptions, and is self-serve from the dashboard; no specific dollar figures or tier table are published in a verifiable form.(unverified) | Managed cloud control plane, with a Sync integration backing config, files, and memory to the cloud. | Built-in cross-channel memory: the Sync integration backs files, conversations, and memory to the cloud so context persists across channels. | Native | Voice (Twilio + Discord), Discord, web/mobile chat, SMS + calls via Twilio, Google Workspace (Meet/Calendar/Drive), Microsoft Teams, Slack, and Sync. | Nine model providers behind one proxy per Alfe's site — Anthropic, OpenAI, Gemini, DeepSeek, Mistral, xAI, OpenRouter, Zhipu, and MiniMax — with per-tenant policy and BYOK overrides. | No |
| LangGraph PlatformFramework | 4.3 | Framework is free (MIT). Managed: Developer $0/seat (up to 5k base traces/mo), Plus $39/seat/mo, Enterprise custom; usage add-ons include deployment runs at $0.005/run. | Open-source framework self-hosted on your own infra; managed cloud deployment available; self-hosted and hybrid deployment gated to the Enterprise tier. | Durable execution across restarts and cold starts, persistent checkpoints (payloads up to 25 MB), and semantic search for long-term memory, with per-subagent state isolation via threads. | Native | Docs cite 1000+ integrations; tools are consumed as built-ins plus remote (HTTP) and local MCP servers via langchain-mcp-adapters. | Many via langchain-<provider> packages: OpenAI, Anthropic, Google (Vertex + GenAI), AWS, Ollama, Groq, Azure AI, Mistral, xAI, Cohere, DeepSeek, and more. | Yes |
| CrewAIFramework | 4.0 | Basic is free (visual editor, AI copilot, GitHub integration, 50 workflow executions/mo). Enterprise is custom-priced (request a trial); no public figure. | Hybrid — the open-source library is self-hosted via CLI; the Enterprise console offers managed deployment on CrewAI or private infrastructure. | A unified Memory class where the LLM infers scope and importance on save; retrieval is ranked by similarity, recency, and importance, defaulting to LanceDB with custom backends supported. | Native | Connectors including Gmail, Google Drive, Outlook, Teams, OneDrive, HubSpot, Slack, and Salesforce, plus tools to call existing CrewAI automations or Amazon Bedrock Agents. | Native SDKs for OpenAI, Anthropic, Google, Azure, AWS Bedrock, and Snowflake Cortex; many more via LiteLLM (Llama, Mistral, Groq, Ollama, Perplexity, OpenRouter, and others). | Yes |
| AutoGPTManaged | 3.2 | Managed cloud: Pro $42.50/mo and Max $272.00/mo (billed annually), with pay-as-you-go credits; no free cloud tier. Self-hosting is free. | Dual — a paid managed cloud, or free self-hosting via Docker Compose. | Blocks include basic operations (memory management, data structures, file I/O) and agents are file-aware; a deeper persistence model is not documented in primary sources.(unverified) | Native | Modular "blocks" the site describes as 45+ platforms — Google Workspace, GitHub, HubSpot, Discord, Reddit, X/Twitter, LinkedIn, Firecrawl, Perplexity, and media-generation providers. | OpenAI, Anthropic (Claude), Google (Gemini), Mistral, Cohere, DeepSeek, xAI (Grok), Perplexity, Meta Llama, Qwen, Groq, Amazon Nova, and Nous Research. | Yes |
| MastraFramework | 4.2 | Self-hosted is free (Apache-2.0). Cloud Starter $0/mo with usage metering, Cloud Teams $250/mo, Enterprise custom; add-ons include memory at $10/1M tokens. | Dual — self-host the open-source TypeScript library on your own infra, or use managed Mastra Cloud. | Message history, working memory (persistent structured user data), semantic recall (embedding-based retrieval), and observational memory; requires a storage provider (PostgreSQL, LibSQL, or Redis). | Native | Tools and MCP servers, plus framework adapters for Next.js, React, Astro, Express, SvelteKit, and Hono, and data-source connectors. | A model router to 90+ providers via a single provider/model API — OpenAI, Anthropic, Google, DeepSeek, Groq, Mistral, xAI, plus gateways like OpenRouter and any Vercel AI SDK model. | Yes |
| E2BHosting | 3.7 | Hobby (free) includes a one-time $100 usage credit, up to 1-hour sessions, and 20 concurrent sandboxes. Pro is $150/mo + usage (24-hour sessions, up to 100 concurrent). Usage is per-second (e.g. 1 vCPU at $0.000014/s). | Hosted cloud that spins up on-demand secure Linux VM sandboxes reachable both internally and via a sandbox URL; self-hostable given the open-source codebase. | Sandbox persistence — pausing saves both filesystem and memory state, and connect() resumes the identical state, with paused sandboxes kept indefinitely. This is infrastructure state, not agent/LLM memory. | Native | First-party Python and JS/TS SDKs, Templates for defining environments, and MCP tools from a Docker catalog (Browserbase, Stripe, GitHub, Notion, and more). | Model-agnostic — you bring your own LLM; E2B runs the code and tools, not the model. | Yes |
| ModalHosting | 3.5 | Starter is $0 base + $30/mo of free compute credits; Team is $250/mo base + $100/mo credits; Enterprise is custom. Usage is per-second (e.g. H100 GPU at $0.001097/s, CPU at $0.0000131/core/s). | Fully managed, multi-cloud serverless with a proprietary container runtime and scheduler that autoscales from zero to 1000+ GPUs. | Durable primitives — Volumes (a distributed file system persisted across invocations), plus Dicts and Queues. This is infrastructure state, not agent memory. | Partial | A Python-native SDK and CLI as the primary interface, plus Sandboxes, Notebooks, web-endpoint deployment, and observability. | Not applicable — pure compute; you bring and run your own models. | No |
| Fly.io MachinesHosting | 3.7 | Usage-based, billed per second while running (e.g. shared-cpu-1x/256MB roughly $2/mo; performance-1x/2GB roughly $32/mo). Volumes are $0.15/GB-month and egress varies by region. No standing free tier is advertised. | Hardware-virtualized micro-VMs on Fly-operated servers across 18 documented regions on six continents. | Durable state via Fly Volumes (persistent block storage) and snapshots; Machines can be persistent or stopped (stopped root filesystem is retained). Machine RAM itself is ephemeral. | Native | A Machines REST API and flyctl CLI, Fly Launch for containerized deploys, and a native fly mcp command surface with client wiring for major editors. | Not applicable — pure compute/hosting; commonly used to host MCP servers and AI workloads. | No(unverified) |
| Cloudflare AgentsFramework | 4.5 | No separate charge for the SDK — you pay for underlying Workers and Durable Objects. Workers has a free tier (100k req/day) and a paid plan from $5/mo; Durable Objects bills requests, duration, and (from Jan 2026) SQLite storage. | Cloudflare serverless — Workers for compute with each agent as a Durable Object; deploy once and Cloudflare runs the agents across its global network. | A core strength — each agent is a Durable Object with SQLite-backed local storage, persistent memory, durable identity, and recoverable execution, with no external database required. | Native | Decorator methods, state sync, scheduling, WebSockets, AI chat persistence, workflow orchestration, email/voice, browser agents, sandboxed code execution, and React hooks (useAgent, useAgentChat). | Workers AI by default, plus OpenAI, Anthropic, Google Gemini, or any other provider. | Yes |
| Replit AgentIDE agent | 3.5 | An effort/credit model with monthly credits per tier: Starter (free) with daily Agent credits and one published project; Core $20/mo (billed annually) with $25/mo of credits; Pro $95/mo with $100/mo of credits; Enterprise custom. | Runs in Replit's cloud Project Editor / workspace; apps are hosted and shipped via Replit Publish and Deployments, with no local runtime. | Checkpoints let you roll back to a prior state, and multi-artifact projects share one backend/database; Pro adds 28-day database rollbacks. No separate long-term agent-memory feature is documented in primary sources. | Native | Built-in Database and User Auth, plus connectors such as Stripe, OpenAI, BigQuery, Linear, Slack, and Notion, with more via MCP. | Not named in public documentation — docs describe the Agent as powered by large language models and expose effort tiers, but no specific providers or models.(unverified) | No |
| Vercel AI SDKSDK | 3.8 | The SDK is free and open source (MIT); you pay the underlying model/token costs of whichever provider you route to. Vercel's hosting and AI Gateway are billed separately. | A self-deployed library — you install and run it inside your own app or runtime; it is not a managed service. | Provides message/state and streaming primitives; built-in long-term memory and conversation persistence are left to your own store.(unverified) | Native | Tool calling, MCP servers and tools, structured object generation, and multimodal (text, speech, transcription, image, video) generation. | Multi-provider by design — Anthropic, OpenAI, Google, xAI/Grok, Mistral, Meta, Perplexity, DeepSeek, and more (16+ providers, 100+ models). | Yes |
| OpenAI Agents SDKSDK | 3.8 | The SDK is free and open source (MIT); you pay OpenAI (or the chosen provider's) API/token costs. | A self-deployed library (Python, with a TypeScript version too) that runs in your own process and infrastructure. | Sessions provide automatic conversation-history management across agent runs — a persistent working-context layer. | Native | Built-in tools (web search, file search, computer use via the Responses API), function tools, MCP servers, agent handoffs, and tracing/observability. | OpenAI-first (uses the Responses API by default for OpenAI models) but provider-agnostic — supports 100+ other LLMs via LiteLLM and any-llm adapters. | Yes |
| Claude Agent SDKSDK | 4.2 | The SDK library is free (the Python repo is MIT), but use is governed by Anthropic's Commercial Terms of Service and you pay Claude API token costs (or via Bedrock/Vertex/Azure). | A self-deployed library that runs the agent loop inside your own process; a separate hosted option (Managed Agents, an Anthropic-run sandbox per session) also exists. | Sessions maintain context across exchanges (files read, analysis, history), are resumable and forkable, and are stored as JSONL on your filesystem; it also supports CLAUDE.md memory files and automatic context compaction. | Native | Built-in tools (Read, Write, Edit, Bash, Glob, Grep, WebSearch, WebFetch), lifecycle hooks, subagents, permissions, Skills, slash commands, and plugins. | Anthropic-first (Claude models); authenticates via the Anthropic API, Amazon Bedrock, Google Vertex AI, or Microsoft Azure AI Foundry. No non-Claude model support is indicated. | Yes |
| DifyFramework | 3.7 | Dify Cloud: Sandbox is free (200 message credits, 1 workspace, 5 apps); Professional is $59/workspace/mo (5,000 credits/mo); Team is $159/workspace/mo (10,000 credits/mo). A self-hosted community edition is available separately. | Managed cloud (dify.ai) or a self-hosted open-source community edition on your own infrastructure. | RAG knowledge pipelines to make your data LLM-ready, plus conversation and app state within workflows. | Native | Agentic workflows, RAG pipelines, bidirectional native MCP (consume tools and publish as an MCP server), a tool/plugin ecosystem, and observability. | Model-agnostic — any major LLM, plus local models via Ollama and any OpenAI-compatible API. | Yes |
| Relevance AIManaged | 4.0 | Usage-based on two units — an "Action" (a tool run) and "Vendor Credits" (model cost). Free sign-up includes a one-time 1,000 Vendor Credits plus 200 Actions/month; top-ups are $80 per 1,000 Actions and $20 per 10,000 Vendor Credits; Enterprise is contact-sales. | Fully-managed cloud SaaS; no self-host option is documented. | A "Knowledge" RAG feature — upload files or sync Google Drive, SharePoint, Notion, and websites to ground agents. | Native | The homepage cites 1,000+ apps (HubSpot, Salesforce, Slack, Gmail, LinkedIn, Apollo, Notion), with Enterprise referencing 2,000+. | Multi-provider routing across Claude (Anthropic), Gemini (Google), and GPT (OpenAI). | No |
| LindyManaged | 2.5 | Plus $49.99/mo, Pro $99.99/mo, Max $199.99/mo, Enterprise custom. No free tier — a 7-day free trial (no card). Metered in credits (most tasks 1-3 credits on basic models, ~10 on large; minimum 1/task; no rollover). | Fully-managed cloud SaaS that connects to third-party apps; no self-host. | Uses personal context plus inbox/calendar context; a standalone long-term memory product is not described in the documentation we reviewed.(unverified) | None | Marketed as 100+ integrations — Gmail, Outlook, Google Calendar, Slack, Notion, HubSpot, Salesforce, Teams, and Zoom. | Multi-model, selectable per agent; the docs reference GPT-4 and Claude, with Claude as the default. | No |
| n8nFramework | 4.2 | Self-hosted Community Edition is free. Cloud (billed annually): Starter EUR20/mo (2.5K executions), Pro EUR50/mo (10K executions), Business EUR667/mo (40K, self-hosted), Enterprise custom — all with unlimited users, workflows, and integrations. | Two options — n8n Cloud (hosted) or self-hosted on your own infrastructure, including a free Community Edition on GitHub. | AI Agent nodes take memory sub-nodes: "Simple Memory" (in-instance, not persisted across sessions), "Postgres Chat Memory" for persistence, and a "Chat Memory Manager" for advanced management. | Native | 400+ integrations per the GitHub repository (the homepage markets 500+); provider-agnostic LLM nodes via LangChain. | LangChain-based; documented chat models include OpenAI, DeepSeek, Google Gemini, Groq, and Azure, and you can connect any model. | No |
| Zapier AgentsManaged | 3.8 | Metered by "activities" (a trigger, knowledge answer, action run, web browse/scrape, or web search each count as 1 activity). Free includes 400 activities/mo; Pro includes 1,500/mo; Enterprise is custom. Agents do not consume the core Zapier task quota. Exact paid-tier dollar figures for Agents were not available at access time (the pricing URL returned 404). | Fully-managed cloud SaaS (no self-host); agents run in-platform, in the background, and via a Chrome extension. | Knowledge sources — add FAQs, docs, and public links the agent draws on (each a billable activity); a distinct long-term conversational memory primitive is not described.(unverified) | Native | 9,000+ connected apps and 30,000+ actions (the latter exposed via Zapier MCP). | The product page references a "fine-tuned AI model" but names no underlying LLM provider.(unverified) | No |
Score is our editorial opinion (0–5), applied by one rubric to every platform including Alfe.
Product and company names are trademarks of their respective owners and are used here only to identify the products under discussion.