Skip to content

context-kit

Context-engineering plugins for GitHub Copilot, APM, and Claude Code — get the right information in front of your agent, and keep the wrong information out.

License GitHub Copilot APM Claude Code Local first

Get started Follow a recipe Review trust boundaries

context-kit is a context-engineering plugin pack for GitHub Copilot CLI, APM (Agent Package Manager), and Claude Code. Its spine is a set of complementary retrieval modalities — lexical, structural, code-intelligence, structured-data, history, semantic (RAG), graph, and durable memory — plus a routing agent that picks and composes them. Default workflows keep indexes and reviewed records on your machine; configured model endpoints, providers, and allowlisted commands can reach external systems. Around that retrieval spine, eleven shipped plugins add orchestration, steering, read-only verification and change-impact analysis, controlled runtime evidence, cross-session handoff, and authoring quality.

Why context-kit

  • Retrieval, done right

    A routing agent picks the cheapest modality that answers the question — exact-token, code-shape, schema-path, git history, or meaning — and composes them: narrow lexically, rerank by vectors, pin exact lines.

    Architecture

  • Local-first semantic search

    local-rag chunks and embeds a corpus with ollama, indexes it with turbovec, and can fuse FTS5/BM25 lexical candidates with vectors using deterministic reciprocal-rank fusion.

    local-rag

  • One source, three hosts

    The same plugins install directly into GitHub Copilot CLI (copilot plugin), Microsoft's APM (apm install), and Claude Code (/plugin) — no manual copying of skill folders.

    Getting started

  • Spend tokens where they count

    plan-execute keeps planning frontier-quality and delegates bulky work to a cheaper executor. context-steering places each rule at the cheapest layer that still fires.

    plan-execute

  • Verify before you trust

    A read-only verifier checks repository claims with file:line evidence, while /analyze-impact maps a proposed change's prospective blast radius without implementing or executing it.

    verify

  • Escalate to runtime evidence

    runtime-evidence runs only exact, pre-reviewed command IDs from a user-owned allowlist when static verification cannot settle a runtime claim, then returns bounded artifacts to verify.

    runtime-evidence

  • Carry context across sessions

    context-handoff manually writes and resumes bounded, verified task state, validating repository identity and freshness before saved claims are trusted.

    context-handoff

  • Recall durable project memory

    memory preserves reviewed evidence, primary memories, cue anchors, freshness, and supersession. An optional MemPalace adapter adds project-isolated provider recall and opt-in lifecycle capture.

    memory

  • Author more like these

    plugin-forge scaffolds portable plugins, keeps manifests in lockstep, and enforces a deterministic catalog discovery-quality budget.

    plugin-forge

Install in seconds

copilot plugin marketplace add mbeacom/context-kit
copilot plugin install code-search@context-kit   # auto-installs retrieval-core

apm marketplace add mbeacom/context-kit
apm install code-search@context-kit   # also pulls retrieval-core

/plugin marketplace add mbeacom/context-kit
/plugin install code-search@context-kit   # auto-installs retrieval-core

See Getting started for the full plugin list, requirements, and your first search.

Operate with clear boundaries

  • Cookbook — six task-oriented journeys that compose retrieval, verification, continuity, and orchestration.
  • Security and trust — decide what can execute, what leaves the process, and where retained data lives.
  • Troubleshooting and lifecycle — verify first run, diagnose refusal modes, update safely, and uninstall.

Pick a modality by what you know

You know… Modality Example
an exact token / regex / filename lexical rg -t py 'def login' · fd -e ts
the code shape, not the text structural sg -p 'logger.debug($$$)' --lang js
the symbol — its defs / refs / callers code-intelligence global -xr parseConfig · ctags -R
a JSON/YAML schema path structured-data jq '.scripts' package.json
when / why code changed history git log -S'retry' -- src/
only the meaning / intent semantic (RAG) rag query "how do we handle backoff" --name notes
the corpus is an Obsidian vault graph obsidian backlinks file="Project X"
a prior decision, constraint, procedure, or episode durable memory /recall-memory "why did we change retries?"

The retrieval-strategist agent (or the retrieval-strategy skill) chooses which one fits — and they compose.