Claude Code

Add ContextCache as an MCP server in Claude Code so terminal sessions can list your caches, search content, and create chunks. The CLI supports adding MCP servers with a single command.

MCP endpoint

Use this URL when registering ContextCache:

https://contextcache-mcp.dylany.workers.dev/mcp

Option A — Add via CLI

Fastest setup

From a terminal, run the command below. Claude Code registers ContextCache as a remote MCP server and handles authentication on first use.

claude mcp add --transport http contextcache https://contextcache-mcp.dylany.workers.dev/mcp

Start or restart Claude Code. The first time the server is used, your browser opens so you can sign in to ContextCache and approve access.
Verify the connection. Ask Claude Code: “List my ContextCache caches using MCP.”
To remove the connection later, run claude mcp remove contextcache.

Option B — Config file

Requires Node.js 18+

If you prefer a file-based configuration, add or merge the snippet below into your Claude Code MCP config:

{
  "mcpServers": {
    "contextcache": {
      "command": "npx",
      "args": [
        "-y",
        "mcp-remote",
        "https://contextcache-mcp.dylany.workers.dev/mcp"
      ]
    }
  }
}

Restart Claude Code after editing the file. For the most current setup steps, refer to the Claude Code MCP documentation.

Recommended: custom instructions

Save the text below into a CLAUDE.md file at the root of your project (or in the Claude Code user settings) so the assistant bootstraps ContextCache automatically at the start of each session.

# ContextCache MCP — Custom Instructions

You have access to ContextCache, a structured knowledge store, via MCP tools. Follow these instructions whenever the ContextCache MCP server is connected.

## Start of every conversation

At the beginning of each new chat, bootstrap your context automatically:

1. Call `list_caches` to see which caches are available.
2. For each relevant cache (or the one the user specifies), call `get_instructions` to load the cache's system prompt — these are standing rules, identity notes, and priorities that should shape every response.
3. If the conversation topic is clear, run a quick `search` against the cache so you start with relevant knowledge instead of guessing.

Do this proactively — don't wait for the user to ask.

## Available tools

| Tool | Purpose |
|---|---|
| `list_caches` | List all caches you have access to |
| `get_instructions` | Load a cache's system prompt (rules, identity, priorities) |
| `search` | Semantic search over a cache — use for recall |
| `list_chunks` | Browse chunks at root level or within a parent |
| `get_chunk` | Read a chunk's full content, children count, and links |
| `create_chunk` | Add a new chunk (defaults to Inbox if no parent given) |
| `update_chunk` | Edit a chunk's title or content (re-embeds automatically) |
| `delete_chunk` | Remove a chunk (cannot delete the system Inbox) |
| `move_chunk` | Reorganize — move a chunk under a different parent |
| `create_link` | Link two chunks (supports, contradicts, expands, summarises, see also, references) |
| `trigger_job` | Queue a background automation (re_embed, cleanup, auto_link, deduplicate, summarize, cluster) |
| `get_job_status` | Check whether a queued job has finished |

## Workflow guidelines

- **Read before write.** Always check existing state (`search`, `list_chunks`, `get_chunk`) before creating or updating content.
- **Confirm before mutating.** When creating, updating, or deleting chunks, state the target cache and summarize the planned change in one sentence before executing.
- **Verify after mutating.** After any create/update/delete, re-read the affected item and report what changed.
- **Use search liberally.** If a user's question could be answered by cache content, search first rather than relying on your own knowledge.
- **Keep Inbox tidy.** New items land in the Inbox by default. If the user has folders, suggest moving items to the right place.

## Safety

- Never delete or overwrite content unless the user explicitly asks.
- If cache choice is ambiguous, ask one clarifying question.
- Never invent cache IDs or chunk IDs — always use values returned by tools.
- If a tool call fails, return the exact error text and suggest one concrete fix.

Download as .md file