AI agent frameworks for trading, compared by control and state
An agent framework decides how a workflow keeps state, calls tools, pauses for a person and recovers after failure. It does not supply a brokerage account or make an order safe. OpenClaw and Hermes are operational agent runtimes. LangGraph, CrewAI, Microsoft Agent Framework and Claude Agent SDK are developer frameworks. Fensory is listed separately as a venue layer, not as another framework.
Published 2026-10-18, last edited 2026-10-18. Every competitor claim was read against its own source on 2026-10-18. The page is re-checked when a source changes and at least quarterly. Fensory appears in this comparison and sells a trading account.
The comparison
| Framework or layer | What it is | How it runs | MCP and workflow model | State and memory | Human approval | Trading limitation |
|---|---|---|---|---|---|---|
| OpenClawAgent framework | Self-hosted assistant gateway and persistent agent runtime. | A workstation or server you operate. One gateway is one trust domain. | Native MCP server management, tool policies and chat-driven or scheduled agent workflows. | Stateful sessions plus workspace files and searchable memory. | Tool policies, command approvals and MCP approval modes. Full-permission configurations can run without prompting. | Approval depends on configuration, and the runtime supplies no native order, exposure or wallet limit. |
| Hermes AgentAgent framework | Self-hosted personal agent with tools, skills, memory, subagents and schedules. | Local, Docker, SSH, Daytona, Singularity or Modal backends. | Native MCP configuration, reusable skills, scheduled automations and messaging surfaces. | Bounded MEMORY.md and USER.md stores, conversation search and reusable procedural skills. | Smart, manual and off modes for dangerous commands. Container backends treat the container as the security boundary. | Its command review evaluates shell danger rather than financial intent. It does not replace trading-account controls. |
| LangGraphAgent framework | Low-level orchestration framework for long-running, stateful agents. | Application library, LangSmith Cloud, hybrid, self-hosted or standalone Agent Server. | MCP through the LangChain MCP Adapter, plus explicit graph nodes for validation, approval, execution and recovery. | Thread checkpoints plus durable cross-thread stores, with production database backends available. | Interrupts can pause a graph before execution and resume from a checkpoint after review. | It requires engineering for authentication, idempotency, risk checks, storage and recovery. The current LangChain MCP interface is beta. |
| CrewAIAgent framework | Python framework for role-based agent teams and event-driven Flows. | Self-hosted Python framework or CrewAI AMP managed deployment. | Direct MCP support over stdio, Streamable HTTP and SSE with static or contextual tool filtering. | Persistent Flow state and unified memory with configurable storage backends. | Human feedback can be inserted into Flows, and managed deployments can pause and resume through review webhooks. | Human review is inserted by the developer rather than applied universally. Autonomous Crews are less predictable than explicit Flows. |
| Microsoft Agent FrameworkAgent framework | Microsoft successor to AutoGen and Semantic Kernel for agents and explicit workflows. | Microsoft Foundry Hosted Agents or self-hosting, with durable execution support. | Local and hosted MCP tools over stdio or HTTP, plus workflows that can pause for approval or information. | Session state with configurable history, context providers and durable workflow extensions. | Approval-required tools and request-for-information workflow patterns can pause execution. | Python self-hosting packages remain prerelease. AutoGen projects should be evaluated as migrations, not treated as the current recommendation. |
| Claude Agent SDKAgent framework | Python and TypeScript SDK for embedding the Claude agent loop in an application. | A Python or TypeScript process operated by the developer. Managed Agents is a separate Anthropic product. | MCP over stdio, HTTP, SSE or in-process servers, with allow and deny rules and tool callbacks. | Disk-backed sessions with continue, resume and fork. Cross-machine use needs an external SessionStore. | Permission modes, canUseTool callbacks and pre-tool hooks can pause or reject actions. | It is Claude-specific, sessions do not replace external trading state, and the Python package currently declares alpha status. |
| FensoryOursVenue layer | Hosted trading account and MCP venue layer, not an agent framework. | Remote Streamable HTTP reached through OAuth. | The MCP server exposes account, market, perpetual futures and Hyperliquid spot tools through supported clients. | Account and order state live in the product. Agent reasoning and workflow state remain with the connected client or framework. | a kill-switch, an access expiry, a market allow-list and a per-order cap, all enforced on the account. An order above the cap is rejected rather than trimmed | Fensory does not turn a framework into a safe strategy. The workflow still needs fresh-data checks, idempotency, review and recovery appropriate to the use case. |
Choose the runtime for the job, then add the trading account
OpenClaw and Hermes give an operator a persistent assistant with messaging, memory, schedules and tools. They are useful when the goal is an agent a person can reach from chat and keep running on controlled infrastructure. Their approval settings govern whether a tool call proceeds, but neither runtime knows that an order is too large, a market is disallowed or a price is stale unless the connected account or workflow supplies those checks.
The developer frameworks trade convenience for control. LangGraph is strongest when the team wants an explicit state machine with checkpoints, validation nodes and interrupts before execution. CrewAI is useful for role-based research teams, with Flows providing a more predictable execution path than an autonomous Crew. Microsoft Agent Framework fits Azure and .NET teams that need managed hosting and durable workflows. Claude Agent SDK provides the shortest Claude-native path to MCP, sessions, hooks and permission callbacks.
None is a universal winner because the relevant choice is the operating model. A solo operator who wants a persistent assistant has a different problem from a trading team building a deterministic workflow with audit requirements. Fensory sits after that choice: the framework decides when and why to call a tool, while the account and venue decide what the call may do.
The runtimes, frameworks and venue layer
OpenClaw
Official projectA user-operated gateway that connects messaging channels to persistent agents, memory, tools, schedules and local or remote MCP servers.
- What it reaches
- Whatever the connected tools expose.
- Cost
- Open-source runtime. Infrastructure and connected-service costs apply.
Choose this if: You want an always-on personal operator reachable from chat and are willing to operate and secure its gateway.
Worth knowing: Approval depends on configuration, and the runtime supplies no native order, exposure or wallet limit.
Hermes Agent
Official projectA Nous Research runtime that can run locally or in several isolated backends and connect to MCP servers with tool filtering.
- What it reaches
- Whatever the connected tools expose.
- Cost
- Open-source runtime. Model, infrastructure and connected-service costs apply.
Choose this if: You want a self-improving personal runtime with MCP, schedules, memory and messaging, and can govern the skills it changes.
Worth knowing: Its command review evaluates shell danger rather than financial intent. It does not replace trading-account controls.
LangGraph
Official projectA graph-based runtime combining deterministic code with model-driven steps, durable checkpoints and explicit interrupts.
- What it reaches
- Whatever the application and connected tools expose.
- Cost
- Open-source framework with optional managed deployment costs.
Choose this if: You need explicit order-validation nodes, checkpoints, retries and a reviewable state machine around execution.
Worth knowing: It requires engineering for authentication, idempotency, risk checks, storage and recovery. The current LangChain MCP interface is beta.
CrewAI
Official projectA framework that separates autonomous Crews from structured Flows. For trading execution, Flows provide the more explicit sequence and state model.
- What it reaches
- Whatever the connected tools expose.
- Cost
- Open-source framework with optional managed-platform costs.
Choose this if: You want role-based research agents and a structured Flow for the steps that may reach execution.
Worth knowing: Human review is inserted by the developer rather than applied universally. Autonomous Crews are less predictable than explicit Flows.
Microsoft Agent Framework
Official projectA framework combining agents, state, middleware, telemetry and durable workflows with Azure-oriented managed hosting options.
- What it reaches
- Whatever the connected tools expose.
- Cost
- Framework packages plus Azure or self-hosted infrastructure costs.
Choose this if: Your team already operates on Azure or .NET and needs managed identity, telemetry and durable enterprise workflows.
Worth knowing: Python self-hosting packages remain prerelease. AutoGen projects should be evaluated as migrations, not treated as the current recommendation.
Claude Agent SDK
Official projectAnthropic libraries for Claude tools, sessions, hooks, context management and direct MCP connections.
- What it reaches
- Whatever the connected tools expose.
- Cost
- SDK plus model, infrastructure and connected-service costs.
Choose this if: You want a Claude-native application with direct MCP integration and the existing Claude agent loop.
Worth knowing: It is Claude-specific, sessions do not replace external trading state, and the Python package currently declares alpha status.
Fensory
OursOfficial projectFensory supplies a remote trading MCP server and account controls. A client or framework decides when to call it and how the broader workflow keeps state.
- What it reaches
- MCP: Hyperliquid spot and perpetual futures. App: tokenized US equities and treasuries, and gold
- Cost
- Free to connect. 0.01% to 0.05% on perpetual futures and Hyperliquid spot, nothing on tokenized equities, yields or bridging
Choose this if: You have already chosen the agent runtime and need a hosted trading account with controls enforced outside the model.
Worth knowing: Fensory does not turn a framework into a safe strategy. The workflow still needs fresh-data checks, idempotency, review and recovery appropriate to the use case.
How we checked
Every framework row was checked against official documentation on 18 October 2026. MCP support means the framework can connect to MCP servers in general. It does not mean Fensory has verified that framework as an end-to-end client.
OpenClaw and Hermes are grouped as operational runtimes because they ship an agent a person can run and reach. LangGraph, CrewAI, Microsoft Agent Framework and Claude Agent SDK are grouped as developer frameworks because a team builds the application around them.
The comparison treats human approval as a workflow control rather than a trading limit. A prompt, hook or approval callback can decide whether a tool call runs. Only the connected account or venue can enforce a market allow-list, order cap, withdrawal scope or revocation at the trading boundary.
Fensory publishes this page and appears last as a venue layer. It is not scored as a framework and no framework is described as a verified Fensory client without end-to-end evidence.
Questions people ask
- Which AI agent framework is best for trading?
- There is no universal winner. LangGraph offers the most explicit workflow control. Microsoft Agent Framework fits Azure and .NET teams. CrewAI fits role-based research teams. Claude Agent SDK is the shortest Claude-native path. OpenClaw and Hermes are better when the goal is a persistent personal agent rather than a custom application.
- Do these frameworks connect to MCP servers?
- All six have an official MCP path. Their transports, authentication and maturity differ. General MCP support does not prove that a specific remote server has been tested end to end in that framework.
- Does a human approval step make an AI trading agent safe?
- It can stop an action before it runs, but it does not validate the market, size, price or account state unless the workflow checks those facts. Account-enforced limits, credential scope, stale-data checks, idempotency and audit logs remain separate controls.
- What is the difference between an agent runtime and a venue layer?
- The runtime holds the conversation, memory, schedule and workflow. The venue layer exposes market and account tools and processes permitted orders. A production system normally needs both, with a clear boundary between reasoning and financial authority.
- Can I run a trading agent without managing a server?
- A hosted MCP endpoint removes the need to operate the trading server. The agent runtime may still need a workstation, managed deployment or cloud process, depending on the framework. Fensory Cloud agents are a separate future product and are not assumed on this page.