Getting started

context-kit is a plugin marketplace that three hosts install from directly — the same plugins, from a single Markdown source, with no manual copying of skill folders.

Host	Register the marketplace	Install a plugin
GitHub Copilot CLI	copilot plugin marketplace add mbeacom/context-kit	copilot plugin install <name>@context-kit
APM	apm marketplace add mbeacom/context-kit	apm install <name>@context-kit
Claude Code	/plugin marketplace add mbeacom/context-kit	/plugin install <name>@context-kit

Installing code-search or verify automatically pulls in the retrieval-core spine. runtime-evidence and context-handoff depend on verify, so either pulls verify and then retrieval-core transitively. memory depends on context-handoff, so it pulls the complete continuity and verification chain.

Install the plugins

Pick the host you use. Install one entry plugin for the journey you need rather than the entire catalog. code-search is a good first choice and brings the retrieval spine with it.

GitHub CopilotAPMClaude Code

copilot plugin marketplace add mbeacom/context-kit
copilot plugin install code-search@context-kit

See the GitHub Copilot guide for host configuration and the lifecycle guide for verification, updates, and uninstall.

apm marketplace add mbeacom/context-kit
apm install code-search@context-kit

APM adds a project lockfile, dependency resolution, target deployment, and audit/policy checks. See the APM guide.

/plugin marketplace add mbeacom/context-kit
/plugin install code-search@context-kit

Run /reload-plugins, then inspect /plugin → Installed or claude plugin list --json.

Choose an entry plugin:

Journey	Install	Also needed at runtime
Search code, data, or docs	code-search	rg; optional search CLIs
Search a corpus semantically	indexkit	Ollama, embedding model
Narrow an Obsidian vault, then rerank	obsidian + indexkit	Obsidian CLI or rg/fd
Verify repository claims	verify	—
Verify, then observe runtime behavior	runtime-evidence	Runner path: Python 3, POSIX, reviewed allowlist. Optional-tool path: an approved host tool
Verify, then hand off	context-handoff	Python 3
Recall durable project memory	memory	Python 3.10+; Git; MemPalace optional
Plan, then delegate execution	plan-execute	A host with the required workflow/subagent support
Choose where context belongs	context-steering	—
Author portable plugins	plugin-forge	validation toolchain for maintainers

See Plugins for every component and dependency, or start from a complete cookbook journey.

Requirements

The skills degrade gracefully and tell you what's missing. Default storage is local, but a configured model endpoint, provider, or user-allowlisted command may access a network or other external state. Allowlisting does not prove a command has no side effects.

• code-search

  ---

  Needs rg (ripgrep); everything else is optional. Run bash plugins/code-search/scripts/check-tools.sh from a clone to see what's installed and the brew install … line for the rest.

• indexkit

  ---

  Needs a running ollama with an embedding model (ollama pull nomic-embed-text). Claude Code and GitHub Copilot CLI auto-bootstrap the indexkit CLI; APM and manual users can pip install indexkit or bootstrap it once (below). uv is needed only for that bootstrap.

• obsidian

  ---

  Optional: the official obsidian CLI (with Obsidian running) for graph-accurate queries; otherwise it falls back to rg/fd. Set your vault with CONTEXT_KIT_OBSIDIAN_VAULT.

• runtime and handoff tools

  ---

  runtime-evidence and context-handoff require Python 3 for their bundled scripts, which use only the standard library. The runtime runner also requires POSIX and refuses Windows before execution, and needs a user-owned exact-ID JSON allowlist; the handoff validator is cross-platform. Runtime evidence's optional-tool path ships no runner, so it needs neither Python nor POSIX — only an approved, host-exposed tool. Handoffs default to .context-kit/handoff.md.

• memory

  ---

  Needs Python 3.10+ and Git on PATH for local reviewed records. Optional provider recall uses a separately installed MemPalace CLI (uv tool install mempalace). Automatic capture is disabled by default. Project scope is explicit except for GitHub Copilot Desktop's trusted POSIX per-session binding for canonical github.com/owner/repository origins.

Check your toolbox

Run the bundled checker from a clone of the repository to see which optional CLIs are present and the exact brew install … command for the gaps:

bash plugins/code-search/scripts/check-tools.sh

Complete first-run setup

Verify installation with the host-specific checks in Troubleshooting and lifecycle. Claude Code and GitHub Copilot CLI both run the plugin's SessionStart hook, so indexkit bootstraps itself there. APM does not deploy plugin hooks, so it starts with no venv — install the published CLI and the plugin launcher picks it up from PATH:

pip install indexkit          # or: uv tool install indexkit
ollama pull nomic-embed-text
indexkit list

To build the plugin's own venv from a clone instead — and the route to take when bootstrap.sh --check reports a stale venv, since the launcher runs an existing venv in preference to PATH:

export CONTEXT_KIT_DATA="$HOME/.local/share/context-kit"
bash plugins/indexkit/scripts/bootstrap.sh
export PATH="$PWD/plugins/indexkit/bin:$PATH"
ollama pull nomic-embed-text
indexkit list

Use CONTEXT_KIT_* variables in portable profiles. The GitHub Copilot guide contains the canonical cross-host variable table, and each plugin page documents its own defaults and Claude fallback.

Before configuring a remote Ollama host, runtime allowlist, memory provider, or Claude lifecycle hooks, read Security and trust boundaries.

Run a first journey

Once a plugin is installed, ask your agent naturally — the routing agent loads the right skill for the task:

• "Use the retrieval strategy to find where retry backoff is handled."
• "Use code-search to find structural React useEffect cleanup issues."
• "Use indexkit to query my notes for billing open questions."
• "Use the Obsidian RAG bridge to search notes linked to Project X."
• "Analyze the change impact of renaming this event field."
• "Collect runtime evidence for this unable-to-check health endpoint claim."
• "Write a handoff before I continue this task in another session."
• "Recall why this project changed its retry policy, then verify the source."

The key habit is retrieve, then pin: use search to find candidates, open the primary file, and cite exact lines before relying on the answer. The cookbook walks through that flow plus graph/rerank, verify/observe, verify/handoff, recall/verify, and plan/execute.

Next steps

• Architecture — the modality model and how the plugins compose.
• Cookbook — task-oriented multi-plugin journeys.
• Security — trust boundaries, retained data, hooks, providers, and command execution.
• Troubleshooting — first-run checks, refusal modes, updates, uninstall, and data locations.
• Plugins — a page for each plugin in the catalog.
• Contributing — validate, lint, and test the marketplace.