The choice depends on your stack and where you plan to host.
Orkeon next to six other projects, with the same questions asked of each. Each cell links to the project’s public documentation. Nothing here measures speed, reliability or cost.

When to choose Orkeon, when to prefer another framework
- Choose Orkeon in .NET, for teams declared in a file and a boundary enforced by the runtime, with no required cloud.
- Prefer Microsoft Agent Framework (MAF) if you target Microsoft Foundry, or if per-tool approval must work without turning anything on.
- Use both: an Orkeon team runs as a MAF agent; a MAF agent serves as the brain or a tool of an Orkeon agent.
- Prefer CrewAI or LangChain / LangGraph if your stack is in Python.
Orkeon and Microsoft Agent Framework
Orkeon does not replace Agent Framework: it plugs into it. What it brings of its own is the boundary and continuity.
The boundary: every file, endpoint and budget an agent may use is declared, then enforced by the runtime. Continuity: the team you tuned is the team you integrate, without a rewrite.
The two comparison tables
A gap in one project is not a point for another. The numbers in brackets link to the sources.
| Criterion | Orkeon | Microsoft Agent Framework | Semantic Kernel |
|---|---|---|---|
| The project | |||
| Vendor, license | One maintainer, MIT; a dedicated legal entity is being set up. | Microsoft, MIT [1]. | Microsoft, MIT; its README says it is succeeded by Agent Framework [2]. |
| Vendor’s cloud offering | None: neither models nor hosting are sold. | Microsoft Foundry; self-hosting is documented [3]. | Azure OpenAI and Foundry through optional connectors [4]. |
| Release status | 1.0.0, released; public API frozen per package, experimental parts marked [Experimental]. |
Stable; hosting, A2A and DevUI in preview [1]. | Supported “for the foreseeable future”; new features go to Agent Framework [2]. |
| Governance | One maintainer; public repository, continuous integration (new tab) and architecture decisions (new tab). | Repository of the microsoft organization; no foundation named [1]. |
Repository of the microsoft organization; no foundation named [2]. |
| Writing a team | |||
| Languages | C# / .NET 10; a team is written in YAML, TypeScript or C#. | .NET and Python; Go in preview [5]. | C#, Python and Java, without orchestration or observability in Java [4]. |
| Model providers | Vendors, aggregators and local models (the list (new tab)), and any OpenAI-compatible endpoint; Ollama proven in continuous integration. | Twelve providers listed, including Ollama; any IChatClient [6]. |
Ten connectors and OpenAI-compatible endpoints; Ollama locally [4]. |
| Team described in a file, run without code | A crew.yaml (agents, tasks, mounts, budget), run by orkeon run. |
“Declarative Agents” in YAML; running without code is not described in the pages read [7]. | Declarative specification in Python, “experimental” [8]. |
| Validation without a model call | orkeon run crew.yaml --validate: VALIDATION OK, zero tokens (reference (new tab)). |
Nothing equivalent in the pages read. | Nothing equivalent in the pages read. |
| No-code interface | Orkeon Studio: Windows desktop; two terminal applications for Linux, macOS and Windows. | DevUI, a sample application for testing and debugging, in preview [17]. | None in the pages read. |
| The boundary | |||
| Tool execution boundary | Virtual file system with declared mounts. Without a domain list, the web tools reach the whole public web. Shell confinement, turned off in a container. | Per-tool approval; FIDES (information flow control), in Python, “experimental”. No documented virtual file system [9]. | Function invocation filters; no documented file system boundary [10]. |
| Budget and resumption | Autonomous mode: the budget the team declares, with five dimensions; without a declared budget, neither tokens nor duration are capped. Other modes: token and cost cap set by the host, none by default. Checkpoints are opt-in. | Iteration limit; checkpoints. No documented token or cost budget [11]. | Maximum iterations and turns; tokens measured, not capped [12]. |
| Log per tool call | One ToolExecution event per call, without the arguments; OTLP export (security (new tab)). |
One execute_tool trace per call; the arguments if sensitive data is enabled [13]. |
One activity per function; no outcome event described [14]. |
| Integrating and delivering | |||
| Protocols | MCP, client and server (the server over stdio), and A2A: experimental. Microsoft Agent Framework in both directions. | MCP client and server; A2A and AG-UI (a protocol between agent and user interface) in preview [15]. | MCP client, documented in Python; A2A absent from the pages read [16]. |
| Delivery forms | NuGet, dotnet tool, apt package (signed source), unsigned .msi installer, macOS archive not notarized, container image, orkeon-host service. |
NuGet and PyPI; ASP.NET Core hosting in preview; container to Azure [3]. | NuGet, PyPI, Maven [4]. |
CrewAI is a team framework in Python, LangChain / LangGraph a library and a graph engine, OpenClaw and Hermes Agent two personal assistants. This table takes eleven criteria from the first one, under the same labels, and adds one on human approval.
Scroll the table from left to right.
| Criterion | Orkeon | CrewAI | LangChain / LangGraph | OpenClaw | Hermes Agent |
|---|---|---|---|---|---|
| Vendor, license | A framework for agent teams. MIT; one maintainer, a dedicated legal entity is being set up. | A framework for agent teams. MIT, crewAI, Inc. [18] | A library and an agent graph engine. MIT; the langgraph-api server is under the Elastic-2.0 license [21][23] |
A self-hosted personal assistant: a gateway and messaging channels, not a team framework. MIT, OpenClaw Foundation [24] | A “self-improving” personal agent, not a team framework. MIT, “Copyright (c) 2025 Nous Research” [26] |
| Languages | C# / .NET 10; a team is written in YAML, TypeScript or C#. | Python, 3.10 to 3.13 [18] | Python and JavaScript / TypeScript [21] | TypeScript, Node 24 [24] | Python 3.11 and later; Electron desktop application [26] |
| Release status | 1.0.0, published on the releases page (new tab) | 1.x series, published on PyPI [18] | langgraph and langchain in the 1.x series, published on PyPI [21] | Date-numbered versions, published on npm [24] | 0.x series, published on GitHub; PyPI is several versions behind [26] |
| Model providers | Vendors, aggregators and local models (the list (new tab)), and any OpenAI-compatible endpoint; the quickstart example runs on Ollama in continuous integration. | 20 providers listed, the others through LiteLLM; Ollama, LM Studio; OpenAI-compatible endpoint [19] | “1000+ integrations”, with no count of providers; Ollama; OpenAI-compatible base_url [22] |
74 provider entries, including 7 routers or gateways and 8 local runtimes (Ollama, LM Studio, vLLM…); api: openai-completions [25] |
43 entries and “any OpenAI-compatible API endpoint”; Ollama, LM Studio, vLLM; Nous Portal recommended [27] |
| Team described in a file, run without code | A crew.yaml run by orkeon run. |
Yes: agents/*.jsonc and crew.jsonc, loaded by crewai run; the former Python + YAML scheme remains available [18][19] |
Not in open source: langgraph.json points to Python or JS code; no-code through LangSmith Fleet, proprietary [22][23] |
Agents and channels declared in openclaw.json (JSON5), identity in SOUL.md and AGENTS.md; one assistant per agent, no declared team [25] |
config.yaml, SOUL.md, skills in Markdown; subagents started by the model (delegate_task), no declared team [27] |
| Validation without a model call | orkeon run crew.yaml --validate: VALIDATION OK, zero tokens. |
Nothing in the pages read; crewai test runs the team [19] |
compile() checks the graph structure (orphan nodes); no dedicated command [22] |
openclaw doctor --lint, static, read-only [25] |
hermes doctor, hermes config check; connectivity is a GET /v1/models [27] |
| Tool execution boundary | Virtual file system with declared mounts, Access denied refusal, code sandbox, Roslyn analyzer. Without a domain list, the web tools reach the whole public web. Shell confinement, turned off in a container. |
No native sandbox: code execution is deprecated, and the docs point to E2B or Modal, both remote [19] | Deep Agents: by default, a virtual file system in the LangGraph state (StateBackend); optional local shell “with no sandboxing” (LocalShellBackend); isolation through sandbox backends (LangSmith, AgentCore, Daytona…) [22] |
Sandbox “off by default”: Docker, Podman, SSH; file access bounded to a root [25] | local terminal by default, “with no isolation”; Docker, SSH, Modal optional; writes bounded by HERMES_WRITE_SAFE_ROOT [27] |
| Human approval, command allowlist | Permission gate, opt-in: the request goes to the host’s approval channel; without a channel, the call is refused. Allowlisted shell, confinement turned off in a container [configuration (new tab)]. | PRE_TOOL_CALL hooks with a request to the human; no command allowlist in the pages read [19] |
interrupt() and HumanInTheLoopMiddleware per tool; no command allowlist [22] |
Per-command approval (ask: always), per-agent allowlist of patterns; tools.allow / deny [25] |
Approval of dangerous commands (smart / manual / off), command_allowlist, a denylist that cannot be bypassed [27] |
| Budget and resumption | Autonomous mode: the budget the team declares, with five dimensions; without a declared budget, neither tokens nor duration are capped. Other modes: token and cost cap set by the host, none by default. Checkpoints are opt-in. | max_iter, max_rpm, max_execution_time; tokens counted, not capped; resumption through @persist and crewai replay [19] |
recursion_limit, limits on model and tool calls; no cost budget; SQLite or Postgres checkpoints, rollback [22] |
48-hour timeout per attempt; no token or cost limit in the pages read; sessions in SQLite, no checkpoints [25] | max_turns (500); no token or cost budget; SQLite sessions, /resume, /rollback through a shadow git repository [27] |
| Log per tool call | One ToolExecution event per call, without the arguments; OTLP export, which the .NET Aspire dashboard reads. |
OpenTelemetry spans to CrewAI AMP; third-party integrations [19] | LangSmith, the vendor’s service; OTLP to a collector with LANGSMITH_OTEL_ONLY [23] |
diagnostics-otel plugin: one openclaw.tool.execution span per call, OTLP/HTTP [25] |
Native OTLP limited to gateway health; per-tool-call traces through the NeMo Relay integration, the bundled Langfuse plugin (opt-in) or a community OTLP plugin [27] |
| Protocols | MCP client and server (the server over stdio) and A2A: experimental, [Experimental]. |
MCP client; A2A client and server [19] | MCP client (langchain.mcp namespace in beta) and MCP server through Agent Server; A2A server through Agent Server (/a2a/{assistant_id}) [22][23] |
MCP client and server; A2A v1.0 client and server, through a plugin [25] | MCP client and server (plugin); A2A v1.0 through a plugin; ACP [27] |
| No-code interface | Orkeon Studio (Windows desktop; terminal on Linux, macOS and Windows), in its own repository. No commercial offering. | Crew Studio, proprietary; CrewAI AMP, free tier and Enterprise on quote [20] | LangSmith Studio (LangSmith key required), Fleet; $0 and $39 per seat plans, Enterprise; self-hosting under an Enterprise license [23] | Control UI (web), macOS, Windows and mobile applications; “no paid tier, hosted service, or token” [25] | TUI, web dashboard, Hermes Desktop; Nous Portal: subscriptions from $0 to $200 per month, Hermes Cloud [28] |
Documentation consulted in November 2026. Orkeon: version 1.0.0 (new tab).
Documentation
- The CLI and validation (new tab)/reference/cli
- Security: virtual file system, Guardian, log (new tab)/architecture/security
- The tool inventory of version 1.0.0 (new tab)github.com
- APIs marked
[Experimental]and their identifiers (new tab)/reference/experimental-apis - Known limitations (new tab)/reference/limitations