MCP Context Engine
AgentPack exposes local repo context through CLI and MCP workflows. The MCP path lets an AI coding agent request task-specific context instead of relying on one large static prompt.
Use AgentPack as an MCP context engine when an agent needs fresh local repository context for a concrete task.
Core MCP flow
readiness()proves the current host can call AgentPack MCP tools and returns version/tool status.start_task(task)writes the task and creates a ranked context pack.get_context()returns the latest pack and refreshes once when the task or repo snapshot changed.route_task(task)returns a read-only route: relevant files, rules, skills, commands, safety warnings, and an agent prompt.get_skill(name_or_path)loads one recommended skill'sSKILL.mdcontent on demand.
This keeps the first prompt smaller while still letting the agent retrieve more detail when needed.
Local-first design
AgentPack does not need cloud indexing, hosted LLM calls, embeddings, or a vector database for scan, summarize, rank, pack, stats, or benchmark. It uses repo-local signals such as paths, symbols, imports, tests, git changes, summaries, and configured rules.
Try the router
agentpack route --task "fix billing webhook retry handling"
Use JSON output when wiring results into scripts:
agentpack route --task "fix billing webhook retry handling" --format json
Debug skill routing directly:
agentpack skills recommend --task "fix billing webhook retry handling" --explain
Why MCP helps
MCP is useful when agents otherwise spend turns searching, opening files, and refreshing stale context manually. AgentPack keeps that selection step explicit, local, and measurable.
MCP also lets agents fetch context after the working tree changes. Static markdown can go stale; AgentPack includes freshness metadata and can refresh when task text, git state, or repo snapshots drift.
What AgentPack can return
An MCP route or context response can include:
- selected files and render modes
- omitted files and selection receipts
- likely tests and suggested commands
- repo-local rules and skills
- safety warnings
- token stats and freshness metadata
The coding agent still owns source inspection, edits, and verification.