Home/guide/MCP Server: Local and Remote Setup

MCP Server: Local and Remote Setupnote

deploymentmcpgetting-started
2 min read · Edit on Pyrite

Pyrite's MCP server lets AI agents interact with your knowledge bases. It can run locally alongside your AI client, or serve a remote Pyrite instance.

Local MCP server (most common)

After installing Pyrite and creating a KB:

```bash pyrite mcp # write tier (default) pyrite mcp --tier read # read-only for untrusted agents pyrite mcp --tier admin # full access including reindex ```

The MCP server runs as a stdio process — your AI client (Claude, Codex, Gemini) launches it automatically via the MCP config. You don't need to start it manually.

Access tiers

| Tier | Tools available | Use case | |------|----------------|----------| | read | search, get, list, backlinks, recent | Untrusted agents, browsing | | write | All read + create, update, delete, link | Your own agents | | admin | All write + reindex, schema, validate | Maintenance tasks |

Rate limiting

The MCP server includes per-client rate limiting to prevent runaway agents:

  • Read tier: 120 requests/minute
  • Write tier: 60 requests/minute
  • Admin tier: 30 requests/minute

    • Available MCP tools

    Read tier:

  • `kb_search` — keyword, semantic, or hybrid search
  • `kb_get` — read a single entry by ID
  • `kb_batch_read` — read multiple entries in one call
  • `kb_list_entries` — browse entries with filters
  • `kb_recent` — recently changed entries
  • `kb_backlinks` — find what links to an entry

    • Write tier (adds):

  • `kb_create` — create a new entry
  • `kb_update` — update an existing entry
  • `kb_delete` — delete an entry
  • `kb_link` — create a link between entries

    • Admin tier (adds):

  • `kb_reindex` — rebuild the search index
  • `kb_schema` — view KB schema and type definitions
  • `kb_validate` — validate entries against schema

    • Connecting to a remote Pyrite instance

    If you're running Pyrite on a server (like the demo at demo.pyrite.wiki), the MCP server currently requires a local install pointed at the same KB data. Remote MCP over HTTP/SSE is on the roadmap.

    Current pattern for remote KBs: 1. Clone the KB repo: `git clone my-remote-kb` 2. Register locally: `pyrite kb add my-remote-kb --name remote` 3. Index: `pyrite index build remote` 4. Connect via MCP as normal

    Changes sync through git push/pull.