reddit-research

A personal-use, retrieval-augmented research tool that answers questions using Reddit discussions as evidence. It searches for relevant threads via a search API, fetches thread/comment data through the official Reddit Data API, ranks the most useful evidence, and returns structured context (optionally synthesized into a cited answer).

It is intended for interactive question answering and short-lived local research. It does not train models on Reddit data, bulk-archive Reddit, or redistribute collected datasets. See the spec for the full design and non-goals.

Install

uv sync                 # core CLI
uv sync --extra mcp     # + MCP server
uv sync --extra dev     # + test tooling

Related MCP server: gemini-grounding

Configure

Configuration splits cleanly in two:

• Secrets → environment / .env (never the TOML). Copy .env.example:

  # Reddit creds — only for the "praw" backend (see Fetch backends below)
  REDDIT_CLIENT_ID=...
  REDDIT_CLIENT_SECRET=...
  BRAVE_SEARCH_API_KEY=...        # or TAVILY_API_KEY (search always needs a key)
  ANTHROPIC_API_KEY=...           # only if you enable synthesis

• Non-secret behavior → reddit-research.toml (see reddit-research.example.toml): backend, user_agent, search/synth provider, comment limits, cache TTLs, etc.

A .env in the current directory (or ~/.config/reddit-research/.env) is loaded automatically. Precedence is real env var > .env file > TOML > default, so any TOML value can be overridden by its env var when needed — but each setting has one canonical home to avoid duplication.

Fetch backends

Reddit gated self-serve API app creation in 2026 (the "Responsible Builder Policy"), so an approved OAuth app is no longer guaranteed. The fetcher therefore supports several backends, selectable per-command with --backend or via [reddit] backend / REDDIT_RESEARCH_BACKEND:

reddit-research evidence "..."                 # auto -> arctic (keyless)
reddit-research answer   "..." --backend praw  # live data, needs an approved app

Compliance note: arctic/json let you run without an approved app, but using them to sidestep API approval sits in tension with the spec's "don't circumvent access controls" non-goal. Intended for genuine personal, low-volume use.

CLI

reddit-research search   "best backup strategy for homelab"
reddit-research fetch     "https://www.reddit.com/r/selfhosted/comments/..."
reddit-research evidence  "what do selfhosted users recommend for backups?" -r selfhosted -r datahoarder
reddit-research answer    "what do Reddit users recommend for homelab backups?"
reddit-research cache stats
reddit-research cache purge --older-than-days 30

Useful flags: --subreddit/-r (repeatable), --limit-threads, --max-comments, --sort, --since-days, --format json|markdown, --no-llm, --show-queries, --verbose.

MCP server

Exposes search_reddit_threads, fetch_reddit_thread, rank_reddit_evidence, and the high-level answer_from_reddit. Two transports (REDDIT_RESEARCH_MCP_TRANSPORT):

stdio (default) — local use; the client spawns the process:

{
  "mcpServers": {
    "reddit-research": {
      "command": "reddit-research-mcp",
      "env": { "BRAVE_SEARCH_API_KEY": "..." }
    }
  }
}

http / streamable-http — a long-lived network server for container/NAS deployment (see below). A bearer token is required; the server refuses to start over HTTP without REDDIT_RESEARCH_MCP_TOKEN, and every request except GET /healthz must send Authorization: Bearer <token>.

Deploy on a NAS (Docker)

The included Dockerfile + docker-compose.yml run the MCP server over HTTP.

1. Configure — create .env next to the compose file:

   BRAVE_SEARCH_API_KEY=...                 # search needs a key
   REDDIT_RESEARCH_MCP_TOKEN=$(openssl rand -hex 32)   # required bearer token
   # ANTHROPIC_API_KEY=...                  # only if you enable synthesis

2. Run — build from source:

   docker compose up -d --build
   curl http://127.0.0.1:8000/healthz            # -> ok

   To deploy a prebuilt image from your own registry instead, set REDDIT_RESEARCH_IMAGE in .env and pull:

   export REGISTRY=your-registry.example.com
   echo "REDDIT_RESEARCH_IMAGE=$REGISTRY/reddit-research-mcp:latest" >> .env
   docker login "$REGISTRY"                      # once
   docker compose pull && docker compose up -d

   To publish a new image after code changes:

   docker build --provenance=false \
     -t "$REGISTRY/reddit-research-mcp:latest" .
   docker push "$REGISTRY/reddit-research-mcp:latest"

   The SQLite cache persists in ./data. The container binds to 127.0.0.1:8000 by default, so it's reachable only through the NAS's reverse proxy (change the ports: mapping to 8000:8000 to expose it on the LAN instead).

3. Reverse proxy (Synology) — Control Panel → Login Portal → Advanced → Reverse Proxy → Create:

   • Source: https://reddit-mcp.<your-domain> (port 443, HTTPS — enables TLS)

   • Destination: http://localhost:8000

   • Enable HSTS as desired; the streamable-HTTP transport streams responses, so leave response buffering off (the default reverse-proxy behavior is fine).

4. Connect from Claude Code:

   claude mcp add --transport http reddit-research \
     https://reddit-mcp.<your-domain>/mcp \
     --header "Authorization: Bearer <your REDDIT_RESEARCH_MCP_TOKEN>"

Security notes: the token gates outbound calls that spend your API keys, so keep it secret and prefer a random 32-byte value. The compose file sets the keyless arctic backend (the public .json path is fingerprint-blocked by Reddit for non-browser clients — see Fetch backends). Switch to praw only if you have an approved OAuth app and set its credentials in .env.

Architecture

question -> query planner -> search provider -> reddit URL extractor
         -> Reddit API fetcher -> cache -> evidence ranker
         -> structured evidence -> optional answer synthesis

Package layout under src/reddit_research/:

Develop

uv run pytest

Status

MVP implemented: search/fetch/evidence/answer CLI, Brave + Tavily providers, three fetch backends (praw/json/arctic), SQLite cache, lexical ranking, JSON/Markdown output, and the MCP server. Later enhancements (async fetching, semantic reranking, branch summarization) are tracked in the spec.