Claude Code Project Context: Beyond CLAUDE.md
CLAUDE.md gives Claude Code stable, repo-specific rules to read at the start of a session. It is a static file, though: it does not update itself, it lives in one checkout, and it is not shared with a teammate's session or another agent. A hosted MCP context layer adds the part CLAUDE.md cannot: decisions, active work, and code impact that change day to day.
What CLAUDE.md already does well
Claude Code reads CLAUDE.md at the start of a session, or Claude Code can use it as a compatibility layer alongside AGENTS.md as the canonical cross-agent instruction file. Either way, it solves a real problem: instead of re-explaining project conventions in every prompt, a developer writes them once and Claude Code starts from them automatically.
That makes CLAUDE.md a good fit for the things that rarely change — code style, folder structure, testing conventions, deployment rules — and a poor fit for anything that changes as often as the work itself.
Where CLAUDE.md stops being enough
CLAUDE.md is text a person wrote. It does not update itself when a decision changes, a task finishes, or the architecture shifts.
It lives in the repository a developer has checked out. A second machine, a second worktree, or a CI job only sees it if the file is there too.
If a teammate's Claude Code session, or a different client entirely, is working in the same repo, CLAUDE.md does not tell either one what the other is doing right now.
CLAUDE.md can describe conventions, but it cannot tell an agent which files a change is likely to affect, or what already broke in a similar change.
CLAUDE.md, session memory, and a hosted context layer
| Layer | Scope | Updates | Shared across agents | Good for |
|---|---|---|---|---|
| CLAUDE.md | Repository checkout, read by Claude Code | Manual edits only | No — each checkout needs its own copy | Stable conventions: style rules, architecture notes, house rules |
| Claude Code session memory | Single chat session | N/A — lost at session end or compaction | No | The current task, inside the current conversation |
| Hosted MCP context layer | Project, reachable by any connected MCP client | Reviewed captures, not automatic scraping | Yes — Claude Code, Cursor, Codex, and other MCP clients query the same state | Decisions, active work, code impact, verification evidence |
Before and after: starting a task in Claude Code
"Implement OAuth." Claude Code reads CLAUDE.md for conventions, then starts searching the codebase, reconstructing architecture, and asking for context a teammate's earlier session already worked out.
Claude Code opens a Start Work Brief through hosted MCP: prior decisions, active work, architecture, likely code impact, and what still needs verification, before the first edit.
Both paths still use CLAUDE.md for stable conventions. The difference is what fills the gap CLAUDE.md was never meant to cover — what changed, what is in flight, and what a change is likely to touch.
How Snipara approaches this
Snipara is a context and memory layer for AI coding agents. It gives each repository a Project Brain powered by Project Intelligence, connecting decisions, active work, code impact, verification evidence and handoffs across Claude Code, Cursor, Codex, Gemini CLI and other MCP clients.
For Claude Code specifically: the primary setup is direct hosted MCP with a project slug and API key, added with claude mcp add. Claude Code can combine that with CLAUDE.md and AGENTS.md, and existing copies of either file are not overwritten by setup — generated copies are kept under .snipara/templates for merge review. Claude Code is also the verified full-hook baseline in Snipara's automation layer, so session-start and tool-use hooks are available where the generated setup installs them.
Limitations
- Claude Code still does all the reasoning and editing — Snipara only supplies context.
- Project Intelligence briefs are reviewable evidence views, not canonical project truth or autonomous approval.
- It does not remove the need to run tests or review diffs before merging.
- Memory updates come from reviewed captures, not automatic scraping of every session.
- Snipara is not an LLM provider — you bring your own Claude Code client and model.
FAQ
Is CLAUDE.md still worth keeping?
Yes. CLAUDE.md is a good place for stable, rarely-changing conventions that a human wrote down deliberately. It is a static file, though, so it will not tell Claude Code what changed this week or what a teammate's session already decided.
Does Claude Code support MCP natively?
Yes. Claude Code has native MCP support, so it can connect to a hosted MCP endpoint the same way any other MCP-compatible client does, without a separate plugin.
Does Snipara replace CLAUDE.md?
No. Snipara's setup does not overwrite an existing CLAUDE.md or AGENTS.md file. Generated copies are kept under .snipara/templates for a human or agent to merge deliberately, so both can coexist.
Does Claude Code get hooks with Snipara?
Claude Code is the verified full-hook baseline for Snipara's automation layer, so lifecycle hooks (session start, pre/post tool use, commit capture) are available where the generated setup installs them. Hosted MCP remains the normal path either way.
How is this different from a longer context window?
A longer context window still starts empty at the beginning of a new session. Project context is about what persists between sessions and across tools, not how much a single session can hold at once.