Beam Weaver

Neither. It is Elixir code built on OTP supervision, with agents, graphs, memory and retrieval as native modules and explicit adapters, and it states plainly that it is not affiliated with LangChain. There is no Python sidecar and no hosted control plane.

It is more than a chat adapter. Decision models load through init_decision_model/2 and run through DecisionModel.invoke/3 for standalone Choice, Score and Noul evaluations that return typed results with probabilities, and routing middleware can use them to select which chat model handles a request.

One dependency plus configuration for the providers you actually use. In exchange you get OTP supervision, Ecto and ETS persistence boundaries, telemetry and tracing inside the same runtime as the rest of your app, at the cost of leaving the Python ecosystem.

A ~> 0.1.19 version constraint says early. The README describes checkpoints, interrupts, resumable execution and provider fallback as features, but publishes no stability policy, upgrade notes or production references, so evaluate it the way you would any 0.1 library.

Agents defined by a module DSL or built at runtime with tools, middleware, structured output, memory and HITL interrupts. Graph workflows with reducers, commands, subgraphs, checkpoints, pending writes and durable execution. Deep agents with planning tools, TODO state, virtual filesystems, skills, subagents, async subagents, context engineering and summarization. Retrieval with loaders, splitters, embeddings, vector stores, record managers and indexing. And persistence through ETS and Ecto-backed memory, checkpoints, caches and vector stores.

Checked-in profiles for current provider families plus permissive fallback profiles for future compatible IDs, with explicit provider prefixes when a name would be ambiguous. Inspect the exact set in your checkout with mix beam_weaver.models.profiles. For TypeSafe there are typesafe:jev-1.13.0, typesafe:jev-latest and typesafe:jev-preview, all supporting Choice, Score and Noul. For deterministic tests there are fake chat and embedding models plus replay transports, and the Jev routing example takes --offline to select fixtures explicitly.

Agents, graphs, model calls, tool calls, subagents, retries, token usage, costs, errors, custom fields and run trees. Local traces work with no external service; WeaveScope export is configured once with an endpoint and WEAVESCOPE_API_KEY. At each call site you attach trace identity and put the dimensions you want to filter by in fields, keeping extra non-indexed context in metadata, alongside the name, user_id, thread_id and execution_mode.