For judges

Everything you need in one page. No setup, no key, no account. Built on the CoinMarketCap DEX API for the Build with CMC: API Hackathon — AI Agents & Automation track.

The claim

Every “LP removed” alert stops at the row that fired it. Forwarding Address follows the wallet across every pool of the token — on this chain and on the asset’s other EVM chains — and rewrites the alert: Rebalance, Migration, Partial, or a real Exit. A plain API call shows the removal; only the maker-keyed cross-pool join and the adjudication turn “liquidity gone” into “liquidity moved to Ring Exchange (Ethereum), 99.8% recovered”.

The 30-second path

The judged capability is one stdlib Python file and the endpoint it calls is keyless:

git clone https://github.com/edycutjong/forwarding.git && cd forwarding
python3 scripts/forwarding.py investigate --platform ethereum --address 0x1f9840a85d5af5bf1d1762f925bdaddc4201f984

Expected: the trace of every call as it happens (endpoint, status, ms), then a red ● LP REMOVED line, then the verdict — for UNI, as of 2026-09-18 23:36:32 UTC (live_run.json):

◆ MIGRATION · severity amber
    99.8% recovered — $21,287,255 of $21,330,275
    +$21,287,255 into Ring Exchange (Ethereum) · UNI/WETH (ethereum)   2 min 12 s later · pool now holds $102,239,395

Then the agent surface, in one line:

claude mcp add forwarding -- python3 $PWD/scripts/mcp_server.py

and ask Claude Code “where did the UNI liquidity go?” — it calls where_did_liquidity_go and the tool answers with the verdict and the raw rows. Hand it a just-in-time transaction and it refuses. Real session transcript: docs/proof/mcp_session.md.

1. Nothing else is required. No pip install, no .env, no API key, no signup. One thing that can look like a bug is not one: the endpoint is CoinMarketCap’s shared anonymous tier, rate-limited per IP. A clean run of the pinned path is 27.0 s (p50 of 5 live runs, 11 calls at 2 s spacing; the bare command adds the sweep pages — 14 calls in the run above). If the tier is throttling, the script backs off (15 s, 30 s, 60 s) and says so; if your IP’s quota is exhausted it exits 75 with the way through — wait a minute, or export a free key as CMC_API_KEY (an escape hatch, never a requirement; every receipt here was taken with no key set).
2. Watch the trace. Every call is printed as it happens — that is the event’s “visible evidence of a real API call” on the judged surface, and the same calls are committed with their verbatim responses in docs/proof/.
3. Read the verdict. The two rows it was computed from are printed underneath, with the division written out. Change the token, or hand it --txn and --maker, and it follows that wallet instead.

Full annotated transcript with the receipt: DEMO.md.

Receipt — two live runs of the same removal

The bare command picks the largest qualifying removal in the token’s newest 300 rows, so weeks from now it may pick a newer one; --txn 0x5081d9…dcd495 --maker 0x4f0a…b9a9 pins this one (the two flags together skip the sweep and walk the wallet; --maker alone still sweeps). Both runs are committed: live_run.json is the bare command (2026-09-18T23:36:32Z); hero.json is the pinned invocation seed.py made after its watchlist sweep chose this removal (2026-09-18T23:16:31Z) — its selection_rule field records that choosing, its calls are the pinned path (no sweep pages). Same rows, same verdict.

The other three outcomes, captured by the same published rule: EXIT (LINK, 0% recovered), REBALANCE (LINK, 105% recovered), refused (DAI JIT pair, $229,834).

Reproduce

python3 scripts/forwarding.py investigate --platform ethereum --address 0x1f9840a85d5af5bf1d1762f925bdaddc4201f984   # the hero, live — the bare command
python3 scripts/forwarding.py investigate --platform ethereum --address 0x1f9840a85d5af5bf1d1762f925bdaddc4201f984 --txn 0x5081d91ab6eae50971e0dec5b5f2b03561893a638768ec4773197622dddcd495 --maker 0x4f0aa5900b8292273b2f9a178d5468f8048bb9a9   # the same removal, pinned by txn + wallet: reproduces this card after the sweep has moved on
python3 scripts/forwarding.py investigate --platform ethereum --address 0x514910771af9ca656af840dff83e8264ecf986ca   # LINK
python3 scripts/forwarding.py watch --cycles 1                    # the autonomous loop, one pass over 11 tokens
make test                                                         # 324 offline tests
make test-live                                                    # 10 tests against the real CMC contract
make verify                                                       # replay every receipt, assert I1–I6, check the page
pytest tests/test_property.py --hypothesis-show-statistics        # the 2,000

CI / deterministic replay — this is not the product: make bench replays adjudicate() over the committed receipts. Nothing on the judged path reads them; investigate always runs live.

Why only CoinMarketCap

/v1/dex/liquidity-change/list returns, per liquidity event and with no key, the maker address (m), the side (tp), the USD value (tu), the venue (en/f) and the pair — and it accepts maker= as a server-side filter. That one parameter is the product: the same wallet’s adds in every other pool of the token come back in a single call. Remove CoinMarketCap and reproducing the join needs a multi-chain liquidity-event indexer, a per-DEX pool registry, a cross-chain contract registry and a pool-depth oracle — four systems to replace 6 keyless endpoints.

Honest limitations

• Pool identity is venue + pair, not pool address. Rows carry no pool contract; two fee tiers of one pair on one venue collapse into one identity.
• The window is ±6 h, walked with the cursor because startTime is plan-gated keyless. A wallet that comes back a day later reads as an Exit for that window — stated on the card.
• A wallet is not an entity. Liquidity moved through a second wallet is not followed.
• A wallet that cycles more than once in the window shows a share over 100% — the adds are attributed to the whole window, and the card discloses the other removals.
• The anonymous tier throttles per IP. The CLI backs off and says so; the hosted page’s live box shares one IP across every visitor, so the page leads with the dated receipt.
• The base rate follows wallets on the same chain only — its EXIT count is an upper bound.

Links