Ecosystem Map

May 27, 2026 · View on GitHub

contextweaver is a runtime context-compilation layer. It sits inside an existing agent loop and decides what context the model sees this turn.

This page is the adopter-facing version of Where contextweaver fits. The short version: keep your framework, MCP servers, memory system, retriever, model SDK, and observability stack. Add contextweaver when the prompt itself needs bounded, deterministic context control.

Decision Matrix

If you are evaluating...	contextweaver relationship	You probably need contextweaver when...
Agent frameworks	Complementary	The framework runs the loop, but prompt assembly is still too large or too manual.
MCP / FastMCP	Complementary	The server exposes many tools or returns large payloads that should not all reach the model.
Memory systems	Complementary	Long-lived memory exists, but each turn still needs budgeted selection.
RAG / vector retrieval	Complementary	Retrieved docs compete with tool results, facts, and history for the same prompt budget.
Observability	Instrumented by it	You need diagnostics for prompt construction, not another dashboard.
LLM SDKs	Upstream consumer	You want a `ContextPack.prompt` to hand to OpenAI, Anthropic, Gemini, or local models.

You probably do not need contextweaver if...	Better first step
You have not built an agent loop yet.	Pick an agent framework or write the loop first.
You only need document retrieval.	Start with RAG or a vector database.
You only need durable user memory.	Start with Mem0, Zep, LangMem, or your own memory store.
You only need traces, dashboards, or cost reporting.	Start with OpenTelemetry or your existing observability stack.
Your tool catalog and tool outputs are tiny.	Keep the simple prompt path until context pressure appears.

Agent Frameworks

Frameworks such as LangGraph, CrewAI, Pydantic AI, OpenAI Agents SDK, Google ADK, and custom loops own control flow:

when to call the model
when to call a tool
how to retry
when the task is finished

contextweaver owns a narrower question:

Given the current phase and budget, which events, facts, tool cards, and tool
results should enter the prompt?

Use both when your framework has a working loop but still leaves you hand- rolling truncation, tool shortlist prompts, or large-result summaries.

Concrete paths:

Runtime	Start here
LangChain / LangGraph	LangChain + LangGraph guide
OpenAI Agents SDK	OpenAI Agents SDK guide
Google ADK / Vertex AI	Google ADK guide
CrewAI	CrewAI guide
Pipecat	Pipecat guide

MCP and FastMCP

MCP exposes tools and data. It does not decide which of 60, 600, or 6,000 tool definitions should be shown to a model on a specific turn, and it does not make large tool results prompt-safe by itself.

contextweaver can sit in front of MCP-style tools in two ways:

Pattern	What happens
Adapter + routing-only	Convert tool definitions to `SelectableItem`s, then route to a small `ChoiceCard` set.
Gateway/proxy runtime	Expose `tool_browse`, `tool_execute`, and `tool_view` meta-tools over MCP-shaped payloads.

Start with the MCP integration guide, the FastMCP cookbook recipe, or the MCP Context Gateway architecture.

Memory Systems

Memory systems persist knowledge across sessions. contextweaver decides what from memory, history, facts, and tool results earns space in the current prompt.

System	Primary job	contextweaver job
Mem0	Extract and retrieve durable memories.	Budget retrieved memories alongside current events.
Zep / Graphiti	Maintain temporal knowledge graphs.	Pull relevant facts into a phase-specific prompt.
LangMem	Structure semantic, episodic, and procedural memory for LangGraph.	Compile selected memories with tool context and dependency chains.
SQLite / custom stores	Persist events and artifacts.	Query, score, filter, firewall, and render.

See External Memory Backends for the protocol shape. Issue #195 tracks the remaining first-class memory adapters.

RAG and Vector Retrieval

RAG retrieves documents. contextweaver compiles agent context.

They compose cleanly:

Run RAG against your document corpus.
Ingest retrieved chunks as ContextItem(kind=ItemKind.doc_snippet, ...).
Let contextweaver score those chunks against facts, tool results, and recent history under the same token budget.

This is useful when retrieved documents are not the only thing the model needs. For example, an answer may need a document chunk, the tool result that produced an incident ID, and the parent tool call that explains where that result came from.

Observability

contextweaver is not an observability product. It does emit useful diagnostics:

Diagnostic	What it explains
`BuildStats`	Included, dropped, deduplicated, and token usage counts.
Route traces	How routing expanded the graph and ranked candidates.
Context explanations	Per-candidate scoring and drop reasons when requested.
OpenTelemetry integration	GenAI semantic-convention attributes for existing backends.

Use OpenTelemetry or your existing dashboard stack to store and visualize these signals across production traffic.

Positioning Boundaries

contextweaver claims:

bounded tool-card routing
artifact-backed context firewalling
phase-specific prompt compilation
deterministic, LLM-free core execution
inspectable build and routing diagnostics

contextweaver does not claim:

to replace your framework, MCP servers, memory layer, RAG stack, model SDK, or observability system
to guarantee answer-quality improvements
to guarantee a fixed cost reduction for every workload
to make lexical retrieval sufficient for every large catalog