> ## Documentation Index > Fetch the complete documentation index at: https://docs-omnicoreagent.omnirexfloralabs.com/llms.txt > Use this file to discover all available pages before exploring further. # How It Compares > OmniCoreAgent next to the OpenAI Agents SDK, LangGraph, Pydantic AI, the Claude Agent SDK and CrewAI, capability by capability, each cell sourced # How It Compares Checked on **2026-09-26**, against each project's own documentation and repository. Every cell below links to its source in *Sources*. These projects move fast, so a cell can go stale. If one is wrong, please [open an issue](https://github.com/omnirexflora-labs/omnicoreagent/issues) and we will correct it. **Yes** means documented and built in. **Partial** means part of it, a separate package, or a paid product (the notes say which). **No** means the docs say or show it is not there. **Not documented** means we could not find it in the official docs, which is not the same as No. ## The short version **What OmniCoreAgent does differently.** It puts one policy over every action an agent takes: each tool, each command, the network, each MCP server, delegation, background runs. When the policy asks, the run pauses for a person and can resume days later in another process. A budget prices each model call before making it, and when it runs out the run waits for a person instead of failing. Commands run in a sandbox whose network is off by default and *checked* to be off. Every run is recorded locally, with no service to sign up for. The same package also schedules runs and serves them over HTTP. All of this is opt-in: governance is off until you turn it on. **Where the others are stronger:** * **Explicit workflow graphs.** LangGraph is built around them, with per-step checkpoints and "time travel" to earlier state. CrewAI has Flows, and Pydantic AI has pydantic-graph. OmniCoreAgent has no graph API: you compose runs in Python, or give an agent sub-agents. * **Sandbox breadth.** The OpenAI Agents SDK supports local, Docker and seven hosted sandbox providers. OmniCoreAgent has five hosted providers, Docker and local. The Claude Agent SDK isolates commands at the operating system level with no container at all. * **Durable-execution engines.** Pydantic AI and the OpenAI Agents SDK integrate with Temporal, DBOS, Restate and others. OmniCoreAgent keeps its own durable run record instead. * **Hosted tracing and evaluation.** LangSmith, Logfire, OpenAI Traces and CrewAI AMP each come from the same vendor as the framework. OmniCoreAgent exports to OTLP, LangSmith, Opik and JSONL, but has no hosted product. * **TypeScript.** The OpenAI Agents SDK, LangGraph and the Claude Agent SDK have TypeScript versions. OmniCoreAgent is Python only. * **Ecosystem size.** LangChain's integrations and community are far larger. ## Capability by capability | | OmniCoreAgent 0.4 | OpenAI Agents SDK | LangGraph | Pydantic AI | Claude Agent SDK | CrewAI | | ------------------------------------------------------------------------------- | --------------------------------------------------------- | ---------------------------------------------------- | ------------------------------------------------------------------------ | -------------------------------------------- | -------------------------------------------------------------- | --------------------------------------- | | **1. One policy that allows, asks or denies each action** | Yes (opt-in) | Partial: per-tool approval and guardrails | Partial: per-tool approval (`langchain` middleware) | Partial: a guard you write (Harness package) | Yes: rules, modes, hooks | Partial: hooks you write | | **2. A person's approval pauses the run; it resumes later, in another process** | Yes (durable store) | Yes: `RunState` | Yes: `interrupt()` + checkpointer | Yes: deferred tools | Yes: `defer` hook (one tool call per turn) | Yes: Flows `@human_feedback` | | **3. Dollar budget enforced before a model call** | Yes: holds the call's price first; pauses for a person | Not documented (tokens and turns) | No in open source; LangSmith gateway (paid, beta) | Yes: `cost_limit` ends the run once exceeded | Partial: `max_budget_usd` stops after the call that crosses it | No: paid Cost Limit only emails | | **4. Built-in sandboxed command execution** | Yes: Docker, E2B, Modal, Daytona, Vercel, your own, local | Yes: local, Docker, 7 hosted | Partial: Docker or Codex policy (`langchain`); providers via Deep Agents | Partial: Modal, Monty (Harness) | Yes: OS-level, opt-in | Partial: E2B and Daytona tools | | **5. Sandbox network off by default** | Yes, and checked on E2B and Daytona | No: Docker keeps its default network | Partial: yes for the Docker policy; the default runs on the host | Not documented | Partial: sandbox opt-in; new hosts are asked about | No (E2B tool) | | **6. A run survives a process crash and resumes** | Yes (durable store) | Partial: save state yourself, or Temporal and others | Yes: checkpoints | Yes: Temporal, DBOS, Prefect, Restate... | Partial: resumes the conversation | Yes: `checkpoint=True` | | **7. Every run recorded locally, no service needed** | Yes, on by default | Partial: default exporter sends to OpenAI | Partial: checkpoint history; tracing needs LangSmith | Partial: opt-in | Yes: JSONL transcripts | Partial: opt-in log files | | **8. OpenTelemetry export** | Yes (OTLP, LangSmith, Opik) | Not documented | Yes (via `langsmith`) | Yes | Yes (traces in beta) | Yes | | **9. MCP client, local and remote** | Yes (+ OAuth) | Yes | Yes (`langchain`) | Yes | Yes | Yes | | **10. Sub-agents / delegation** | Yes | Yes: handoffs, agents as tools | Yes | Yes | Yes | Yes | | **11. Scheduled / background runs built in** | Yes: cron, interval, durable task store | Not documented | Partial: LangSmith Deployment | Partial: via Prefect and others | Partial: session-scoped scheduling | Not documented | | **12. HTTP server built in** | Yes (OmniServe) | Not documented | Partial: dev server; production is LangSmith Deployment | Partial: dev web UI; adapters for your app | No | Partial: AG-UI package; paid hosted API | | **13. Model providers** | Many (LiteLLM) | Many (OpenAI default; adapters in beta) | Many | Many | Claude only, on several clouds | Many | | **14. Workflow graphs** | No: compose in Python | Partial: Python orchestration | Yes: the core of it | Yes: pydantic-graph | Partial: Workflow tool | Yes: Flows | | **15. Hosted tracing / evaluation from the same vendor** | No | Yes: OpenAI Traces | Yes: LangSmith | Yes: Logfire | Not documented | Yes: AMP | | **16. Languages** | Python | Python; TypeScript (separate) | Python, TypeScript | Python | Python, TypeScript | Python | | **17. License** | MIT | MIT | MIT (library) | MIT | Commercial terms (Python repo MIT) | MIT | ## Sources Read on 2026-09-26: only each project's own docs and repository, no third-party comparisons. The method is in the repository, in `engineering/architecture/comparison-plan.md`. **OmniCoreAgent:** this documentation. [Policy](/docs/reference/policy) (1), [Durable runs](/docs/core-concepts/durable-runs) (2, 6), [Budgets](/docs/reference/budgets) (3), [Sandbox providers](/docs/how-to-guides/sandbox-providers) (4, 5), [Observability](/docs/how-to-guides/observability) (7, 8, 15), [MCP](/docs/core-concepts/mcp) (9), [Sub-agents](/docs/core-concepts/sub-agents) (10), [Background agents](/docs/core-concepts/background-agents) (11), [OmniServe](/docs/how-to-guides/omniserve) (12), [Models](/docs/how-to-guides/models) (13), [Upgrading](/docs/upgrading) (14). **OpenAI Agents SDK** (openai-agents 0.22.3): [guardrails](https://openai.github.io/openai-agents-python/guardrails/), [human in the loop](https://openai.github.io/openai-agents-python/human_in_the_loop/) (1, 2); [usage](https://openai.github.io/openai-agents-python/usage/) (3); [sandbox clients](https://openai.github.io/openai-agents-python/sandbox/clients/) (4, 5); [running agents](https://openai.github.io/openai-agents-python/running_agents/) (6, 12); [tracing](https://openai.github.io/openai-agents-python/tracing/) (7, 8, 11, 15); [MCP](https://openai.github.io/openai-agents-python/mcp/) (9); [handoffs](https://openai.github.io/openai-agents-python/handoffs/) (10); [models](https://openai.github.io/openai-agents-python/models/) (13); [multi-agent](https://openai.github.io/openai-agents-python/multi_agent/) (14); [repository](https://github.com/openai/openai-agents-python) (16, 17). **LangGraph:** [human in the loop](https://docs.langchain.com/oss/python/langchain/human-in-the-loop) (1); [interrupts](https://docs.langchain.com/oss/python/langgraph/interrupts) (2); [built-in middleware](https://docs.langchain.com/oss/python/langchain/middleware/built-in), [LLM gateway spend policies](https://docs.langchain.com/langsmith/llm-gateway-spend-policies) (3, 4); [execution policies](https://github.com/langchain-ai/langchain/blob/master/libs/langchain_v1/langchain/agents/middleware/_execution.py) (5); [checkpointers](https://docs.langchain.com/oss/python/langgraph/checkpointers) (6, 7); [OpenTelemetry](https://docs.langchain.com/langsmith/trace-with-opentelemetry) (8); [MCP](https://docs.langchain.com/oss/python/langchain/mcp) (9); [subgraphs](https://docs.langchain.com/oss/python/langgraph/use-subgraphs) (10); [cron jobs](https://docs.langchain.com/langsmith/cron-jobs) (11); [local server](https://docs.langchain.com/oss/python/langgraph/local-server) (12); [models](https://docs.langchain.com/oss/python/langchain/models) (13); [graph API](https://docs.langchain.com/oss/python/langgraph/graph-api) (14); [observability](https://docs.langchain.com/oss/python/langgraph/observability) (15); [repository](https://github.com/langchain-ai/langgraph) (16, 17). **Pydantic AI:** [guardrails (Harness)](https://pydantic.dev/docs/ai/harness/guardrails/) (1); [deferred tools](https://pydantic.dev/docs/ai/tools-toolsets/deferred-tools/) (2); [agents: usage limits](https://pydantic.dev/docs/ai/core-concepts/agent/) (3); [Modal sandbox](https://pydantic.dev/docs/ai/harness/modal-sandbox/), [code mode](https://pydantic.dev/docs/ai/harness/code-mode/) (4, 5); [durable execution](https://pydantic.dev/docs/ai/capabilities/durable_execution/overview/) (6); [step persistence](https://pydantic.dev/docs/ai/harness/step-persistence/) (7); [Logfire and OpenTelemetry](https://pydantic.dev/docs/ai/integrations/logfire/) (8, 15); [MCP client](https://pydantic.dev/docs/ai/mcp/client/) (9); [multi-agent](https://pydantic.dev/docs/ai/guides/multi-agent-applications/) (10); [Prefect](https://pydantic.dev/docs/ai/capabilities/durable_execution/prefect/) (11); [web UI](https://pydantic.dev/docs/ai/guides/web/) (12); [models](https://pydantic.dev/docs/ai/models/overview/) (13); [graphs](https://pydantic.dev/docs/ai/graph/graph/) (14); [repository](https://github.com/pydantic/pydantic-ai) (16, 17). **Claude Agent SDK:** [permissions](https://code.claude.com/docs/en/agent-sdk/permissions) (1); [user input](https://code.claude.com/docs/en/agent-sdk/user-input) (2); [cost tracking](https://code.claude.com/docs/en/agent-sdk/cost-tracking) (3); [sandboxing](https://code.claude.com/docs/en/sandboxing) (4, 5); [sessions](https://code.claude.com/docs/en/agent-sdk/sessions) (6, 7); [observability](https://code.claude.com/docs/en/agent-sdk/observability) (8, 15); [MCP](https://code.claude.com/docs/en/agent-sdk/mcp) (9); [subagents](https://code.claude.com/docs/en/agent-sdk/subagents) (10); [scheduled tasks](https://code.claude.com/docs/en/scheduled-tasks) (11); [hosting](https://code.claude.com/docs/en/agent-sdk/hosting) (12); [quickstart](https://code.claude.com/docs/en/agent-sdk/quickstart) (13); [TypeScript reference](https://code.claude.com/docs/en/agent-sdk/typescript) (14); [overview and license](https://code.claude.com/docs/en/agent-sdk/overview) (16, 17). **CrewAI:** [tool hooks](https://docs.crewai.com/en/learn/tool-hooks) (1); [human feedback in Flows](https://docs.crewai.com/en/learn/human-feedback-in-flows) (2); [AMP policies](https://docs-platform.crewai.com/platform/en/features/agent-control-plane/policies) (3); [code execution](https://docs.crewai.com/en/tools/ai-ml/codeinterpretertool), [E2B tool](https://github.com/crewAIInc/crewAI/blob/main/lib/crewai-tools/src/crewai_tools/tools/e2b_sandbox_tool/README.md) (4, 5); [checkpointing](https://docs.crewai.com/en/concepts/checkpointing) (6); [crews](https://docs.crewai.com/en/concepts/crews) (7); [tracing](https://docs.crewai.com/en/observability/tracing) (8); [MCP](https://docs.crewai.com/en/mcp/overview) (9); [collaboration](https://docs.crewai.com/en/concepts/collaboration) (10); [automation triggers](https://docs-platform.crewai.com/platform/en/guides/automation-triggers) (11); [frontend](https://docs.crewai.com/en/guides/frontend/overview) (12); [LLMs](https://docs.crewai.com/en/concepts/llms) (13); [Flows](https://docs.crewai.com/en/concepts/flows) (14); [AMP traces](https://docs-platform.crewai.com/platform/en/features/traces) (15); [repository](https://github.com/crewAIInc/crewAI) (16, 17).