Gecko — judgment layer for trading agents
The oracle that
says no.
Execution is commodity. Judgment is scarce.
Gecko sits above any trading agent — Eliza, SendAI, OKX, Backpack, your own. The panel refuses ungrounded calls, cites the canon when it doesn't, and ships eight validated nulls before its first green light.
● V1 advisor in waitlist · MCP-ready
Sits above Eliza · SendAI · OKX · Backpack
First ten builders run free · Solana wallet optional
q: Is the JLP-USDC vault on Kamino worth depositing right now? (entry, 2024-Q3)
0.42 — corpus too thin for directional call
citations · 5
- [1]Kamino vault docsprotocol_native · 0.72
- [2]Howard Marks — On Riskinvestor_canon · 0.41
- [3]Kamino governance forumprotocol_native · 0.39
- [4]Berkshire 2008 letterinvestor_canon · 0.28
- [5]Damodaran — Implied ERPinvestor_canon · 0.22
surviving dissent · 2 voices
> Two voices argue the protocol's incident-free track record offsets the audit gap; the panel could not resolve this against the corpus.
blocker · what would change the answer
- · Will the Kamino protocol secure a credible audit report within the next month?
- · Can the total value locked (TVL) in the vault increase significantly in the coming weeks?
Live envelope shape. Verdict and citations from an actual S24 evaluation run.
Grounded
Cited or it didn't happen
Every verdict ships with the exact chunks that backed it and the dissent that didn't win. If a claim isn't grounded in the corpus, it isn't in the answer.
Calibrated
Defers on thin evidence
When the panel can't find enough to call it, the verdict is defer — with a blocker question naming what would change the answer. No directional call without supporting evidence in the corpus.
Composable
One oracle, any skill
Callable from Claude Code, agent runtimes, CLIs, and whatever you ship next. We're the grounded-knowledge layer that other skills call — not a closed app.
How it works
Four layers above your agent.
- 01
LAYER · JUDGMENT
Oracle
The panel debates the call — retrieving over investor canon (Marks, Damodaran, Berkshire letters) and live protocol data, then returning a verdict envelope: act / pass / defer, calibrated confidence, the citations that drove the call, and the dissent that didn't win. Default behavior is decline when the corpus or chart can't support a directional call.
gecko_trade_research(idea="kamino jlp-usdc entry") - 02
LAYER · SAFETY
Trap node
Every candidate token routes through a honeypot and rug detector (92% catch on 112 adversarial inputs in the current screen) before any execution path opens. If the token can't be sold, the oracle never sees it — the verdict is overridden to pass and the blocker surfaces verbatim.
- 03
LAYER · AUTHORITY
Scoped permissions
Server-wallet pattern: per-session limits, per-venue allowlists, per-asset caps. The verdict becomes a signed authorization the agent uses once — no blanket signing, no shared custody, revocable per-session.
- 04
LAYER · EXECUTION
Swappable adapters
OKX, SendAI, Backpack, Jupiter, your own runtime — slot in behind the same interface. Execution is the commodity layer; swap venues without changing the agent above or re-grounding the oracle below.
Honest state
We defer when the corpus is too thin to call it.
And we tell you what's missing instead of inventing an answer. Every shipped verdict is scored on a rubric that grades citation relevance, hallucination, dissent grounding, and confidence calibration. The numbers are how we know the wedge is real.
From S24 evaluation run · the panel deferred on this query because protocol-specific corpus coverage was below threshold.
q: Is the JLP-USDC vault on Kamino worth depositing right now? (entry, 2024-Q3)
0.42 — corpus too thin for directional call
citations · 5
- …
surviving dissent · 2 voices
> Two voices argue the protocol's incident-free track record offsets the audit gap; the panel could not resolve this against the corpus.
blocker · what would change the answer
- · Will the Kamino protocol secure a credible audit report within the next month?
- · Can the total value locked (TVL) in the vault increase significantly in the coming weeks?
Grounded in
Built on a corpus people actually quote.
Howard Marks · Aswath Damodaran · Berkshire Hathaway letters
Pyth Network · Helius · Solana primary sources
Investor canon, 1977 — present
1.00
verdict accuracy
0.16
brier (calibration)
0.90
defer when thin
Pilot-scale (N=10). Eval JSONs published in the gecko-mcpay-api repo.
FAQ
Common questions.
V1 advisor mode ships across the next two sprints. The first ten builders on the waitlist get founder-onboarded the day it lands; the rest enter cohorts shortly after. Early access is free.
Not in V1. V1 is advisor-mode — verdicts, citations, dissent. Execution adapters (OKX, Backpack, SendAI) ship in a later version behind explicit user opt-in. We don't take custody and we don't claim PnL.
Gecko is a Claude Code skill. Read app.geckovision.tech/skill.md in any Claude Code session and the oracle becomes a tool you can call. Direct API access via api.geckovision.tech is available for early-access cohorts.
Early-access cohorts run free in stub-billing mode. Session pricing for the paid tiers will be announced before V1.0. We won't surprise the waitlist with a price tag.
Two structural differences. First, we retrieve over a categorized corpus of investor canon and live protocol data — not the open web. Second, the panel is built to refuse: when evidence is thin, the verdict is defer with a blocker question, not a confident guess. The rubric grades us on that behavior every release.
Early access
Get on the waitlist.
First ten builders run free. We'll email when V1 ships — no spam.