Radar for context engineers

109 curated tools and processes, grouped by the six jobs of context engineering. Every number links to its source.

Retrieve

Pull the right outside knowledge into context, just in time: search, RAG, embeddings, rerankers, docs and code fetchers.

30 tools
  • BGE-M3
  • Chonkie
  • Chroma
  • Context7
  • +26
Top 3 toolsHide tools
  1. 1.BGE-M3Multilingual embedding model that does dense, sparse and multi-vector retrieval in one model (BAAI).
  2. 2.ChonkieLightweight, fast chunking library (Python and TypeScript) for splitting text before embedding.
  3. 3.ChromaOpen-source search infrastructure for AI: embed, store and query documents for retrieval.

Memory

Keep what matters outside the window across turns and sessions, and bring it back when needed.

13 tools
  • Basic Memory
  • Beads
  • claude-mem
  • Cognee
  • +9
Top 3 toolsHide tools
  1. 1.Basic MemoryLocal Markdown knowledge base that AI assistants read and write through MCP, so you stop re-explaining projects.
  2. 2.BeadsGit-backed issue tracker built as long-term memory for coding agents.
  3. 3.claude-memCaptures what a coding agent does in a session, compresses it with AI and injects the relevant parts next session.

Tools & MCP

Give agents actions and live data through tools: MCP servers, SDKs, registries and tool design.

17 tools
  • Agent Skills (anthropics/skills)
  • Chrome DevTools MCP
  • Code execution with MCP
  • Composio
  • +13
Top 3 toolsHide tools
  1. 1.Agent Skills (anthropics/skills)Anthropic's public repository of Agent Skills: instruction folders an agent loads only when a task needs them.
  2. 2.Chrome DevTools MCPGives coding agents Chrome DevTools: inspect pages, debug and trace performance.
  3. 3.Code execution with MCPExpose MCP tools as code the agent calls from a sandbox, so only the definitions and results it needs enter context.

Compression

Fit more signal into fewer tokens: compaction, summarization, token-efficient formats, repo packers, caching.

13 tools
  • Attention budget
  • code2prompt
  • Compaction
  • Context editing
  • +9
Top 3 toolsHide tools
  1. 1.Attention budgetThe finite attention an LLM spends across its context; every new token draws it down, so context must be curated.
  2. 2.code2promptCLI that converts a codebase into one prompt with a source tree, templates and token counting.
  3. 3.CompactionWhen a conversation nears the window limit, summarize it and restart a fresh window with the summary.

Sub-agents & orchestration

Split work across agents with clean, isolated contexts: agent frameworks, sub-agents, handoffs, workflows.

17 tools
  • Agent Development Kit (ADK)
  • Agno
  • AI SDK (Vercel)
  • BAML
  • +13
Top 3 toolsHide tools
  1. 1.Agent Development Kit (ADK)Google's code-first toolkit for building, evaluating and deploying agents (Python, with a JS SDK).
  2. 2.AgnoFramework and runtime for building, running and managing multi-agent systems.
  3. 3.AI SDK (Vercel)TypeScript toolkit for AI apps and agents: tool calling, streaming and structured output across providers.

Evals & observability

See what went into the context and whether it worked: tracing, evals, regression tests, token and cost tracking.

19 tools
  • Arize Phoenix
  • Braintrust
  • Context clash
  • Context confusion
  • +15
Top 3 toolsHide tools
  1. 1.Arize PhoenixAI observability and evaluation: traces, evals and datasets built on OpenTelemetry.
  2. 2.BraintrustEvals and observability platform (closed source), measured here through its official SDKs.
  3. 3.Context clashWhen new information or tools in the context conflict with other information already there.

Retrieve

30

Pull the right outside knowledge into context, just in time: search, RAG, embeddings, rerankers, docs and code fetchers.

Screenshot of the BGE-M3 website

BGE-M3

Multilingual embedding model that does dense, sparse and multi-vector retrieval in one model (BAAI).

tool

Website captured Oct 19

  1. Prepend chunk-specific context before embedding and BM25 indexing. Anthropic measured 49% fewer failed retrievals, and 67% fewer with reranking.

    Memory

    13

    Keep what matters outside the window across turns and sessions, and bring it back when needed.

    1. Treat files as unlimited, persistent memory: write big observations to disk and keep only paths in context; recite a todo.md to keep goals in focus.

      Tools & MCP

      17

      Give agents actions and live data through tools: MCP servers, SDKs, registries and tool design.

      1. Expose MCP tools as code the agent calls from a sandbox, so only the definitions and results it needs enter context.

        Compression

        13

        Fit more signal into fewer tokens: compaction, summarization, token-efficient formats, repo packers, caching.

        Attention budget

        The finite attention an LLM spends across its context; every new token draws it down, so context must be curated.

        term
          1. When a conversation nears the window limit, summarize it and restart a fresh window with the summary.

            • Automatically clearing stale tool calls and results from the context as it nears its limit (a Claude Developer Platform feature).

              • The share of prompt tokens served from the model's key-value cache; higher means lower latency and cost.

                Sub-agents & orchestration

                17

                Split work across agents with clean, isolated contexts: agent frameworks, sub-agents, handoffs, workflows.

                1. The set of strategies for curating and maintaining the optimal set of tokens during LLM inference, including everything that lands in context outside the prompt.

                  Evals & observability

                  19

                  See what went into the context and whether it worked: tracing, evals, regression tests, token and cost tracking.

                  1. When new information or tools in the context conflict with other information already there.

                    • When superfluous content in the context is used by the model to produce a low-quality response.

                      • When context grows so long that the model over-focuses on it and neglects what it learned in training.

                        • When a hallucination or other error makes it into the context and keeps getting referenced.