contextweaver
Context firewall + tool router for MCP and tool-heavy agents. Phase-specific, budget-aware context engineering with a deterministic core.
Minimal core dependencies · deterministic output · Python ≥ 3.10
Context engineering is the discipline of deciding what goes into a model's context window, when, and at what cost — see the canonical framing.
contextweaver provides two cooperating engines that solve the context window problem for tool-using AI agents:
- Context Engine — eight-stage pipeline: candidates → dependency closure → sensitivity filter → firewall → scoring → dedup → selection → rendering. See the Context Firewall page for the load-bearing firewall primitive and Architecture for the full pipeline.
- Routing Engine — bounded DAG + beam search over large tool
catalogs, producing compact LLM-friendly
ChoiceCards. See the Tool Router page.
Use it when your agent has too many MCP / FastMCP / Python tools, too much tool-result data, or a long history competing for the same prompt budget. Do not use it as a replacement for your agent framework, model SDK, memory database, RAG system, or observability stack.
Get started
10-Minute Quickstart{ .md-button .md-button--primary } Daily Driver{ .md-button } API Reference{ .md-button }
Navigate
| Section | What you'll find |
|---|---|
| Quickstart | Install, first context build, firewall demo, routing demo |
| Daily Driver | When to use the gateway, client instructions, and the operator debug loop |
| Security Model | Gateway data flow, trust boundaries, artifact exposure, and hardening |
| MCP Client Recipes | Claude Desktop, Claude Code, Copilot, and Cursor setup |
| Concepts | Core type glossary: ContextItem, Phase, ChoiceGraph, … |
| Ecosystem Map | How contextweaver compares with agent frameworks, MCP, memory, RAG, and observability |
| Adopter Benchmark Report | Cost, prompt-size, latency, routing-quality, and failure-mode framing |
| Stability | Alpha/Beta/1.0 readiness and public API stability boundaries |
| Launch Kit | Reusable public copy, assets, and responsible-claims checklist |
| Runtime Loop | Four-phase flow diagram and pseudo-code |
| MCP Integration | Tool conversion, session loading, firewall with MCP |
| A2A Integration | Agent cards and multi-agent sessions |
| Architecture | Pipeline details, design rationale, module map |
| API Reference | Auto-generated reference from source docstrings |