Dolphin

You run a Dolphin-based MCP server locally to semantically search code and documentation via the Dolphin API. It starts quickly, supports parallel snippet fetching, and lets you tune performance and reliability for your environment.

How to use

Start the MCP server and connect your MCP client to it. The server runs locally and exposes a standard set of endpoints you can query to retrieve semantically relevant code snippets, with citations. By default, it expects the Dolphin API to be reachable at your configured endpoint. You can adjust how many snippets are fetched in parallel and how long you wait for responses to improve throughput or reduce latency.

How to install

Prerequisites: you need Bun installed on your system. Bun provides the bunx command used to run the MCP server.

Install Bun if you don’t have it yet. Then start the MCP server with Bunx.

# Start the Dolphin MCP server using bunx
bunx dolphin-mcp

Configuration and usage notes

Continue.dev configuration lets you run the MCP server locally and point it at the Dolphin API. You can customize how many snippet fetches happen in parallel and how long you wait for each fetch.

yaml
mcpServers:
  - name: Dolphin-KB
    command: bunx
    args:
      - dolphin-mcp
    env:
      DOLPHIN_API_URL: "http://127.0.0.1:7777"
      # Optional: Performance optimization for parallel snippet fetching
      MAX_CONCURRENT_SNIPPET_FETCH: "8"
      SNIPPET_FETCH_TIMEOUT_MS: "2000"
      SNIPPET_FETCH_RETRY_ATTEMPTS: "1"

Environment and performance presets

Set environment variables to control the Dolphin API endpoint and tuning knobs for parallel snippet fetching.

DOLPHIN_API_URL=http://127.0.0.1:7777
MAX_CONCURRENT_SNIPPET_FETCH=8
SNIPPET_FETCH_TIMEOUT_MS=2000
SNIPPET_FETCH_RETRY_ATTEMPTS=1

Presets for parallel snippet fetching

Choose a preset based on resource availability and desired performance. Each preset sets the three main knobs for snippet fetching.

# Conservative
MAX_CONCURRENT_SNIPPET_FETCH=4
SNIPPET_FETCH_TIMEOUT_MS=1500
SNIPPET_FETCH_RETRY_ATTEMPTS=1

# Recommended
MAX_CONCURRENT_SNIPPET_FETCH=8
SNIPPET_FETCH_TIMEOUT_MS=2000
SNIPPET_FETCH_RETRY_ATTEMPTS=1

# Performance
MAX_CONCURRENT_SNIPPET_FETCH=10
SNIPPET_FETCH_TIMEOUT_MS=3000
SNIPPET_FETCH_RETRY_ATTEMPTS=2

Available tools and actions

The MCP server exposes a set of capabilities that let you search, fetch, and open code with citations. You can filter results by repository, path, and file patterns, then fetch exact code chunks or lines with citations.

Available tools

search_knowledge

Semantically query code and docs across indexed repositories and return ranked snippets with citations.

fetch_chunk

Fetch a chunk by chunk_id and return fenced code with citation.

fetch_lines

Fetch a file slice [start, end] inclusive from disk and return fenced code with citation.

get_vector_store_info

Report namespaces, dims, limits, and approximate counts.

open_in_editor

Compute a vscode://file URI for a repo path and optional position.