arcus-agent-gateway

An MCP (Model Context Protocol) server that gives AI agents read-only, keyless access to market data for the 194 tokenized US equities on Robinhood Chain (Arcus) — quotes, corporate actions, trading capabilities, multipliers and a 13-sector map. No API keys, no auth, no writes: every tool is a GET against the public api.robinhood.com/rhj REST surface, cached and rate-limited so an enthusiastic agent can't hammer the upstream.

Quickstart

Run over stdio (the default, for local agents):

uvx arcus-agent-gateway

Claude Desktop / Cursor config (claude_desktop_config.json or .cursor/mcp.json):

{
  "mcpServers": {
    "arcus": {
      "command": "uvx",
      "args": ["arcus-agent-gateway"]
    }
  }
}

Codex (~/.codex/config.toml):

[mcp_servers.arcus]
command = "uvx"
args = ["arcus-agent-gateway"]

ZCode — register the server and copy the agent skill (all copy-paste):

# 1) start the gateway (keep it running)
uvx arcus-agent-gateway --http --port 8902 &

# 2) register it (merges into ~/.zcode/cli/config.json; workspace .zcode/config.json works too)
python3 - <<'PY'
import json, os
p = os.path.expanduser("~/.zcode/cli/config.json")
os.makedirs(os.path.dirname(p), exist_ok=True)
cfg = json.load(open(p)) if os.path.exists(p) else {}
cfg.setdefault("mcp", {}).setdefault("servers", {})["arcus"] = {
    "type": "http", "url": "http://127.0.0.1:8902/mcp"}
json.dump(cfg, open(p, "w"), indent=2)
print("arcus MCP server registered:", p)
PY

# 3) copy the agent skill (tool guide + watchlist cron recipe)
git clone -q --depth 1 https://github.com/alekskram/arcus-agent-gateway /tmp/aag
cp -r /tmp/aag/.agents/skills/arcus-gateway ~/.zcode/skills/ && rm -rf /tmp/aag
echo "ZCode setup done — restart your session and call any arcus tool"

Hosted form — streamable HTTP on port 8902:

uvx arcus-agent-gateway --http               # 127.0.0.1:8902
curl http://127.0.0.1:8902/health            # -> {"ok": true, "service": "arcus-agent-gateway"}

Tools

All 13 tools are read-only (annotated readOnlyHint: true). Names and parameters are exactly as registered by arcus_mcp/server.py.

Multiplier logic (read this before using prices)

Robinhood Chain tokens carry a multiplier — the corporate-action adjustment factor for the token contract (1.0 = untouched). Splits change it; for example NVDA's 2026-11 split queues pendingMultiplier: "4.0".

• The REST API returns RAW prices. bid/ask from /prices/{symbol} are in token-contract units and are not multiplier-adjusted.
• Adjusted values are computed by this server, never taken from upstream: price_adjusted = round(price_raw × currentMultiplier, 6).
• Raw and adjusted always travel together. Every quote carries bid_raw/ask_raw/spread_raw and bid_adjusted/ask_adjusted/ mid_adjusted next to the multiplier block — never one without the other.
• On-chain quantities (token balances, mint/burn volumes) are natively in adjusted (multiplied) units; REST prices are not. If you compare the two, go through the *_adjusted fields.

Worked example (live fixture, 2026-09-03):

AAPL   currentMultiplier = 1.000566080061092436
       bid_raw   = 327.77   →  bid_adjusted = round(327.77 × 1.000566…, 6) = 327.955544
       ask_raw   = 327.78   →  ask_adjusted = 327.965550
       mid                      mid_adjusted = 327.960547

Pending split warning. When pendingMultiplier is queued (non-empty) and differs from the current one, token_detail() adds a warning like pending split: 1→4.0 on 2026-11-06T00:00:00Z, and quote()'s multiplier block exposes pending + effective_time. After the split lands, raw prices jump by the ratio while *_adjusted fields stay comparable — another reason to always read adjusted values next to the multiplier.

API limits & caching

• Upstream allows 60 req/s without a key; this client self-limits to ≤ 50 req/s (a 20 ms politeness interval between requests, thread-safe).
• Transient failures (429/502/503/504, network errors) are retried up to 3 times with 2s × (attempt+1) backoff.
• Response caches (per process): /assets 5 min, /prices/{symbol} 15 s, /corporate-actions 1 h. market_status() and sector_view() are computed from caches and assets only — they never fan out 194 price requests.

On-chain sources & limits

The v0.2 on-chain tools read two keyless public sources next to the REST API. Both are free, rate-limited and partially restricted — every tool above degrades honestly (per-field omission + warnings[] / error dicts), never with a silent empty answer.

• Public JSON-RPC (default robinhood-rpc.publicnode.com, override with ARCUS_RPC_URL): eth_call (e.g. totalSupply()) works normally. eth_getLogs only answers inside a floating ~45–60-block window behind the latest block — wider or older ranges get HTTP 403 "Archive requests require a personal token" (the backend is Alchemy). The window drifts minute to minute, so transfer_history() walks back in windows that start 48 blocks wide and shrink 48→32→16→8 on each 403, capped at ~14 getLogs requests. eth_getLogs log objects carry blockTimestamp directly — no per-block lookups are needed.
• Fallback RPC (robinhood.drpc.org, ARCUS_RPC_FALLBACK_URL): has no eth_getLogs and no eth_call (JSON-RPC "method not available"); it is used only for eth_chainId / eth_blockNumber.
• Blockscout v2 explorer (robinhoodchain.blockscout.com/api/v2, ARCUS_EXPLORER_URL): requires a browser User-Agent on every request — plain HTTP clients get a Cloudflare 403 "Just a moment…" HTML challenge. Token pages (holders_count, circulating_market_cap, total_supply), one holders page (max 50 rows, no pagination loops) and address token-balances come from here, cached 600 s. token-balances answers in ~0.5 s on plain wallets but hangs 40 s+ on huge contract addresses — the client fails honestly after 15 s with kind explorer-timeout.
• On-chain activity ≠ trades. The chain records Transfer, mint and redeem events between addresses; it knows nothing about order-book trades or prices. Use quote()/quotes() for prices and transfer_history() for token movement.

Raw prices disclaimer

Prices are served exactly as they arrive from Robinhood (RAW) — they are not multiplier-adjusted, and the *_adjusted fields are our computation, not upstream data. All data is for information only, not for trading decisions, and should be verified against the official source before you act on it. No warranty of completeness, accuracy or timeliness.

Optional price history recorder

The Robinhood Chain REST API has no price history endpoint — only current quotes. For the 194 tokenized equities this recorder is the only history source. It is opt-in and disabled by default; nothing is recorded unless you explicitly enable it.

How it works. One tick every 5 minutes (default): fetch a quote for every ACTIVE tradable token through the same rate-limited client (50 req/s cap; average load ≈ 0.65 req/s), append one row per symbol to data/history/snapshots_YYYYMM.parquet (monthly rotation), and maintain a daily OHLCV rollup data/history/daily.parquet (open/high/low/close on mid_adjusted, volume = max of the day's cumulative daily_volume). The rollup runs at the first tick after midnight UTC for the previous day and is idempotent (re-running a day overwrites it, never duplicates).

Enable it:

pip install "arcus-agent-gateway[recorder]"   # adds pyarrow (optional extra)
# systemd (recommended): units ship DISABLED - enabling is your decision
sudo cp deploy/arcus-recorder.* /etc/systemd/system/
sudo systemctl enable --now arcus-recorder.timer   # OnCalendar=*:0/5, Persistent
# or run one tick / a debug loop manually:
python scripts/recorder.py --once
python scripts/recorder.py --limit 5            # debug: first 5 symbols only
ARCUS_INTERVAL_SEC=60 python scripts/recorder.py  # custom interval loop

Data weight & rotation. Full universe (194 symbols) at a 5-minute tick ≈ 2–3 MB/day of snapshots plus ≈ 10 KB/day for the daily rollup. Snapshots rotate monthly (snapshots_YYYYMM.parquet); delete old months when you no longer need raw granularity — daily.parquet is the compact long-term store. Data lands in ~/.local/state/arcus-agent-gateway/history/ (override with ARCUS_GATEWAY_DATA).

Reading it back: the price_history tool serves daily bars and raw snapshots from the same directory. Without pyarrow or data it returns an actionable error pointing here — install the [recorder] extra, never a silent empty answer.

Development

python3 -m venv .venv && .venv/bin/pip install -e '.[dev]'
.venv/bin/pytest -q                      # all offline (fixtures + mocks)
.venv/bin/python scripts/smoke_api_offline.py     # REST client smoke, zero HTTP
.venv/bin/python scripts/smoke_server_offline.py  # full server smoke, zero HTTP

Layout: arcus_mcp/api.py (stdlib REST client), arcus_mcp/server.py (the MCP tools + FastMCP wiring), arcus_mcp/rpc.py (public JSON-RPC client with the adaptive Transfer-log walk-back), arcus_mcp/explorer.py (Blockscout v2 client with the mandatory browser User-Agent), arcus_mcp/sectors.py (validated 13-sector map), arcus_mcp/paths.py (state dir; override with ARCUS_GATEWAY_DATA). Live-captured schema fixtures live in tests/fixtures/ with notes in arcus_mcp/API_NOTES.md. Usage scenarios: examples/use-cases.md.