Findings · public surfaces only · a source on every claim

Seven gaps between GMX's agent story and what an agent actually finds.

Read as holders, not critics: we hold esGMX, we have real forum history, and we applied in the 2026 CEO search. Every finding below is verifiable from public surfaces, one is already fixed and live on our infrastructure, and the rest come with what "fixed" looks like. All of it is GMX's to keep.

The bottom line, in plain English

The door AI agents knock on is not open yet. GMX's own docs tell agents the MCP server "is not yet available," so the integration agents ask for first is a promise, not an endpoint. We built that integration and put it live, free.

The official SDK fails a standard security check on install. Anyone whose build pipeline runs npm audit, a normal, sane default, sees 7 HIGH findings the moment they install GMX's own package. The fix is a one-dependency swap, written out below.

Machines quoting GMX today can get the economics wrong. The March 2026 staking change lives in blog posts, not in the API, so an agent can still tell its user "fees go to stakers" when, right now, they don't.

Where the protocol stands, mid-August 2026

The ground truth the findings sit on, from public trackers and GMX's own releases. Real revenue, real audits, one dominant chain, and a hard year in a market that moved fast.

$360B+

lifetime notional volume, 740,000+ traders, per GMX's own release

cryptoslate.com ↗

$174.9M

V2 Perps TVL today, ~$32M annualized fees, ~$12M protocol revenue

defillama.com ↗

94%

of live TVL sits on one chain (Arbitrum) of five active chains

defillama.com ↗

mcap = FDV

10.45M GMX circulating equals total supply: no emissions overhang

coingecko.com ↗

$90 gate

since 2026-03-04 all staking rewards route to Treasury until GMX trades above $90 (vs ~$6.72 now)

gmxio.substack.com ↗

20 → 1

a public, 20-candidate CEO search; the DAO chose Q, builder of GMTrade, a strong pick by its own process

x.com/GMX_IO ↗

The findings

Ordered by how much each one costs GMX with the people already trying to build on it. Nothing here required anything beyond public pages, public repos, and one clean npm install.

The MCP server agents are told to wait for

Fixed · live

Evidence: "The MCP server is under development and not yet available."
GMX AI Agents documentation, verbatim · docs.gmx.io ↗
Why it matters: GMX's docs court AI agents and ship coding-agent skills for 11 frameworks. But a standing MCP client, the thing an autonomous agent actually connects to, gets "not yet available." The documented plan (read-only phase, then prepare/confirm writes) is exactly the right shape. It just isn't running.
Provisioned: Built and live at vealth.net/mcp since 2026-08-15: five non-custodial tools in precisely the two-phase shape GMX's own roadmap describes: reads free and open, writes as prepare-unsigned / caller-signs / relay-signed. Posted to gov.gmx.io the same day, offered open-source or as a hand-off to whoever picks up the official work. See it answer live →

The official SDK fails npm audit on a clean install

Open · fix drafted

Evidence: Reproduced directly, not read secondhand: a clean npm install @gmx-io/sdk@1.6.4 resolves 304 packages and npm audit returns 16 findings: 9 low, 7 HIGH. The HIGH chain exists to supply a performance.now() polyfill:
@gmx-io/sdk → isomorphic-performance@5.1.1 → react-native-performance@6.0.0 → react-native@0.86.2 // a mobile framework, in a trading SDK → @react-native/virtualized-lists, @react-native/community-cli-plugin → metro / metro-config / metro-transform-worker → image-size // GHSA-w3rx-r6r6-pgpr + GHSA-5p2g-fcmc-qvqq, both CVSS 7.5 fix: replace the isomorphic-performance branch: one dependency swap
Advisories: GHSA-w3rx-r6r6-pgpr ↗ · GHSA-5p2g-fcmc-qvqq ↗
Why it matters: Any organization that gates CI on npm audit fails the build the moment it adopts GMX's own official SDK. The first professional integrator experience is a red pipeline. A Node/browser trading SDK has no React Native use case; the whole chain is dead weight plus two real DoS CVEs.
Provisioned: The exact chain and the one-swap fix, above, free to file as-is. It is also why the live MCP server was built against the documented v2 contracts directly, with zero new dependencies.

Agents can quote economics that stopped being true in March

Open

Evidence: Effective 2026-03-04, all GMX staking rewards route to the Treasury, to be distributed only if GMX trades above $90; the DAO's own proposal language notes the prior buyback approach "did little to move the market." That state lives in Substack posts and forum threads. No field in the API or SDK response carries the current distribution mechanism.
gmxio.substack.com ↗ · cryptopolitan.com ↗
Why it matters: GMX earned its name on "real yield." An agent reading the API today can still tell its user fees flow to stakers, stale by five months. Machine surfaces that let machines say wrong things about your own token are a trust cost paid daily.
Provisioned: A reward-state field in the staking API response (mechanism, effective date, release condition) so no agent can quote the old economics. A one-field change that closes a whole class of misquotes.

Agent onboarding is scattered across five-plus surfaces

Open

Evidence: An integrator has to stitch together the REST API overview, the Oracle API, the GraphQL docs, two SDK clients, the separate gmx-io/gmx-ai skills repo, and the contracts pages. The skills cover coding-agent IDEs; a standing agent gets no single canonical "start here."
docs.gmx.io ↗
Why it matters: Every extra surface an integrator must reconcile is a place two of them disagree. The protocols agents integrate first are the ones where intent maps to one endpoint in one read.
Provisioned: One compressed intent-to-endpoint reference, maintained as a public artifact, the shape our own agent kit already demonstrates for the five live tools.

"Is this pool safe right now" is a dashboard, not a call

Open

Evidence: Chaos Labs runs a real, independent GMX risk portal, for humans. Nothing in the documented API/SDK surface exposes ADL proximity or pool PnL ratios as a queryable score an agent can act on.
chaoslabs.xyz ↗
Why it matters: GM pools are the product: isolated, per-market liquidity an LP chooses deliberately. An agent allocating LP capital cannot read a dashboard; it needs one number with a threshold.
Provisioned: A gmx_get_pool_risk tool: pool, exposure, ADL proximity, one risk score per call. A natural sixth tool for the live server.

Keeper liveness is a documented risk with no health feed

Open

Evidence: V2's two-step model (create order, keeper executes) makes keepers a named liveness dependency: if keepers stop, closes and withdrawals stop with them. Third-party technical writeups document the dependency; no public keeper-health or execution-latency status feed was found.
forgd.com ↗ · docs.gmx.io ↗
Why it matters: Naming a risk in the docs and giving no way to observe it is the worst of both: informed users worry, and nobody can check. A status endpoint is cheap and turns a standing doubt into a number.
Provisioned: A public execution-health endpoint: last keeper execution, median latency, backlog depth. Green most days, and provably green.

No first-party answer to "which chain has depth for this market"

Open

Evidence: Five chains are live, 94% of TVL is on Arbitrum, and the current live-chain footprint is disclosed via third-party trackers. There is no first-party endpoint an agent can call to route an order to the deepest book.
defillama.com ↗
Why it matters: Multichain reach only compounds if a router can see across it. Without a depth-discovery call, four of five chains are invisible to exactly the automated flow the expansion courts.
Provisioned: A cross-chain market-depth endpoint: market, chain, available depth at impact tiers. One call before every routed order.

What the record shows in GMX's favor

The same recon that found the gaps found these, and they are why the gaps are worth closing.

Its worst day was handled with real spine

July 2025: a ~$40M V1 exploit, a fast $5M white-hat negotiation returning ~90%+ of funds, then a completed ~$44M holder compensation program. An uncommonly clean record for an exploit of that size.

halborn.com ↗ · crypto.news ↗

V2's architecture answered the actual failure

Per-market, risk-isolated GM pools directly retired the shared-pool exposure class the V1 exploit lived in. LPs choose their markets; one market's trouble stays in that market.

hackernoon.com ↗

Audit cadence most protocols don't keep

Eight Guardian engagements (88 person-weeks, 365 findings worked), plus ABDK, Certora, Dedaub, and Sherlock reviews, plus a $5M-max ImmuneFi bounty still current as of January 2026.

docs.gmx.io ↗ · immunefi.com ↗

Ahead of nearly everyone on agent groundwork

Coding-agent skills for 11 frameworks, llms.txt doc bundles, and a two-phase MCP plan whose shape is exactly right. The groundwork is real; the last mile is what this engagement provisions.

docs.gmx.io ↗