Batch 29 — Agent Frameworks (TS/Python/JVM SDKs)
Roster (10)
| slug | stars | distribution | cli_binary | local_ui | orchestration | multi_model | tier |
|---|---|---|---|---|---|---|---|
| langchain-mcp-adapters | 3,540 | pip-package | no | none | none | no | A |
| agno | 40,363 | pip-package | unknown | web-dashboard (managed) | hierarchical | yes | A |
| hive-aden | 10,430 | standalone-repo | yes (hive) | web-dashboard (React) | hierarchical | yes | A |
| voltagent | 9,175 | npm-package | yes (create-voltagent-app) | web-dashboard (VoltOps cloud) | hierarchical | yes | A |
| mcp-agent-lastmile | 8,341 | pip-package | yes (mcp-agent) | none | swarm | yes | A |
| yao | 7,541 | standalone-repo | yes (yao) | web-dashboard (SUI) | parallel-fan-out | yes | A |
| cloudflare-agents | 4,984 | npm-package | yes (agents) | web-dashboard (playground) | hierarchical | yes | A |
| docker-agent | 2,962 | standalone-repo | yes (docker agent) | terminal-tui | hierarchical | yes | A |
| nemo-agent-toolkit | 2,333 | pip-package | yes (nat) | web-dashboard (Chat UI) | sequential | yes | A |
| pydantic-ai | 17,322 | pip-package | yes (clai) | web-dashboard (Logfire cloud) | hierarchical | yes | A |
Intra-batch patterns
All 10 frameworks are SDK/library-grade tools operating below the CLAUDE.md/skills layer — they define how agents are programmed, not how Claude Code is augmented. Every framework except langchain-mcp-adapters ships a CLI binary, and seven of ten provide a local UI or cloud dashboard for observability. Multi-model support (routing different models to different roles) is near-universal (9/10), reflecting the maturity of the provider-agnostic abstraction layer (Vercel AI SDK, LiteLLM, PydanticAI's Model protocol). The batch splits along a key architectural axis: some frameworks consume other frameworks' outputs (langchain-mcp-adapters, nemo-agent-toolkit), while most define and run agents independently. Python dominates (6/10) but two Go binaries (yao, docker-agent) occupy unique distribution positions — single binary, zero runtime dependencies.
The most surprising structural finding: mcp-agent-lastmile is the only framework in the entire batch that exposes agents as MCP servers, making agents composable at the protocol level. yao uniquely containerizes Claude Code itself as an agent executor, inverting the normal relationship. cloudflare-agents is the only edge-native (Durable Objects) framework in the catalog.
Most interesting finds
mcp-agent (lastmile): The Swarm pattern (peer-to-peer agent handoffs without a central coordinator) is the only swarm implementation in the catalog, plus zero-API-change transparent Temporal execution. The MCP-as-everything philosophy is the boldest architectural bet in the batch.
Yao: A single Go binary with an embedded V8 engine that containerizes Claude Code itself as an agent executor mode — the only framework where Claude Code is treated as a capability being orchestrated rather than the orchestrator.
Items written as Tier C
None. All 10 frameworks have sufficient public material for full 11-file reports.
Cross-references discovered
- Agno is explicitly referenced by NeMo Agent Toolkit as a supported framework (
nvidia_nat_agnointegration package) — Agno is both a standalone framework and a framework NeMo wraps. - LangChain MCP Adapters is part of the
langchain-aiorganization, making it a first-party bridge between LangChain (seed:spec-kituses langchain internally) and the MCP ecosystem. - Docker Agent builds itself using
./golang_developer.yaml— one of the clearest examples of self-dogfooding in the catalog. - PydanticAI explicitly names LangChain, LlamaIndex, AutoGPT, CrewAI, and Instructor as frameworks that use Pydantic Validation — positioning itself as upstream of most Python agent frameworks.
- VoltAgent uses Vercel AI SDK as its model abstraction layer — the same SDK used by many Next.js-based agent applications, creating an implicit ecosystem bridge.
- Hive (aden-hive) and Claude Flow (seed) are structurally similar (queen+worker, DAG execution, React UI) but Hive auto-generates the DAG while claude-flow requires manual definition.