Skip to content

GeckoVision/gecko-claude

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

51 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

gecko-claude — strategy oracle for autonomous trading agents

Claude Code x402 frames.ag License: MIT

The verdict your trading agent should have called before it lost your money.

A 7-voice adversarial panel — grounded in investor canon (Damodaran, Howard Marks, Berkshire) — returns act / pass / defer with surviving dissent and per-claim citations. Every paid call settles on Solana mainnet via frames.ag. No API keys, no signup, just a wallet.

"I lost money on my last three Solana trades because the LLM just told me what I wanted to hear. So I built the oracle my agent should have called instead." — Ernani, founder


Install (60 seconds, one line)

curl -fsSL https://app.geckovision.tech/install.sh | bash

Then in Claude Code, paste:

Read https://app.geckovision.tech/skill.md and follow the instructions.

Claude walks you through wallet setup (email + OTP, 30s) and funding ($5 USDC covers ~50 verdicts).


What you can do in 90 seconds

1. Ask the panel a single trade question

Should I deposit USDC into the Kamino USDC reserve right now?

Returns a verdict envelope:

verdict:       defer
confidence:    0.7
dissent_count: 2  (surviving — visible on screen)
citations:     [Damodaran "Equity Risk Premium" · Howard Marks Oaktree memo · …]

Every citation links to the actual source — investor canon, not LLM hallucination.

2. Deploy a local advisor agent

Use the gecko-trade-agent skill to deploy my strategy at ~/specs/kamino-dca.json in advisor mode.

Returns an agent_id. The agent lives on your laptop — your keys, your journal — and surfaces opportunities on a schedule. It never signs anything in v0.1.

3. Inspect + fire a fresh verdict

What is my agent doing right now? Then fire a fresh verdict from the panel.

The agent reports heartbeat, journal, and a re-cached panel verdict in ~5 seconds.


Why this is structurally different

Existing surface Sells Picks a side?
Marketplaces (frames.ag, Bazaar, PaySH) Directories of paid agents ❌ Would have to take a side on every listing
LLM chat (ChatGPT, Claude, Perplexity) Conversation ❌ Sycophant by default — tells you what you want to hear
Vector DBs (Pinecone, Weaviate) Retrieval ❌ No verdict; just chunks
Gecko The verdict + dissent + citations + on-chain proof

The strategy-oracle layer is structurally ours because anyone who runs a marketplace cannot also pick winners inside their own catalog.


What's in this repo

Path Purpose
install.sh One-line installer (curl | bash); idempotent, mac+linux.
skill.md Master entry point hosted at app.geckovision.tech/skill.md.
.claude/skills/ 8 skills — 3 trade-vertical (coach / agent / wallet) + 5 V1 (research / ask / sources / extract / fund).
.claude/agents/ 5 sub-agent personas (analyst, validator, architect, builder).
examples/trading-oracle/ Minimal Python client showing a single-shot gecko_trade_research call + envelope parsing.
examples/dex-on-gecko/ Next.js demo app consuming Gecko verdicts in a DEX-style UI.
CLAUDE.md Working agreement that ships into the user's project.
docs/flow.md End-to-end sequence diagram (cold install → first paid verdict).

The tool surface

MCP tool Cost Returns
gecko_trade_research $0.25 (basic, 5 voices) / $0.75 (pro, 7 voices + backtest) Verdict + confidence + citations + surviving dissent. The wedge.
gecko_research $0.10 / $0.75 Idea validation: business plan, market report, V1/V2/V3-scoped PRD. The v1 product — same panel architecture, different vertical.
gecko_ask free Follow-up Q&A grounded in the active session's sources.
gecko_sources free Lists every source with chunk counts.
extract-page ~$0.004/URL Tavily Extract for bot-walled pages.
fund-wallet Walks the user through topping up their frames.ag wallet.

The skills

Trade vertical (v0.1)

Skill Purpose
gecko-trade-coach Conversational strategy builder. Emits a schema-validated spec JSON the agent runtime consumes.
gecko-trade-agent Local advisor runtime. Deploy / list / inspect / pause / reverdict / stop a long-running agent. Advisor mode only in v0.1 — never holds keys, never signs.
gecko-wallet Wallet ops (fund / check balance / receive USDC).

V1 — idea validation

Skill Purpose
gecko-research Paid validation entry point.
gecko-ask Free follow-ups grounded in indexed sources.
gecko-sources Free; lists sources for the active session.
extract-page Paid Tavily Extract (~$0.004/URL) for bot-walled pages.
fund-wallet Wallet topping-up flow.

Chain proof — verify the demo yourself

Every paid call settles on Solana mainnet. A representative settled tx:

4FPSxDGJQykp3j5cbnkGjAd8DVebsHBmazLQNCfEFZ3okKrgWCQi81ujr7aJS8MEbHUDXPEqn7EMAVBdAxwUyWoY

View on Solscan →

Confirmed in 1.6 seconds. Real USDC settlement on mainnet. No API key, no signup.


The numbers

4,874 corpus chunks · 7 panel voices · 80 settled mainnet tx · $0.17 lifetime spend per heavy user
  • Corpus: Damodaran (NYU Stern PDFs), Howard Marks (Oaktree memos), Berkshire (1977–2024 shareholder letters) — all free + public-domain in v0.1.
  • Panel: 7 specialist voices in pro tier (5 in basic) running adversarial debate via AutoGen GroupChat.
  • Settlement: x402 protocol on Solana mainnet via frames.ag wallet — non-custodial, programmatic, ~1.6s confirmation.

How it works

1. You paste one line into Claude Code
   ↓
2. Claude orchestrates: install → wallet OTP → fund prompt → MCP register
   ↓
3. You ask: "Should I deposit USDC into Kamino now?"
   ↓
4. frames.ag wallet pays $0.25 USDC on Solana → gecko-api runs the panel
   ↓
5. Verdict + dissent + citations land in your context. Decide from there.

End-to-end: ~3 minutes from cold install to first verdict. Full sequence in docs/flow.md.


Architecture (three layers, never collapsed)

┌─────────────────────────────────────────────────────────────────┐
│  gecko-trade-coach  →  spec.json (schema-validated)             │
│  (conversational strategy builder)                              │
└─────────────────────────────────────────────────────────────────┘
                              ↓
┌─────────────────────────────────────────────────────────────────┐
│  gecko_trade_research  →  verdict + dissent + citations         │
│  (7-voice oracle; investor-canon grounded; on cadence + trigger)│
└─────────────────────────────────────────────────────────────────┘
                              ↓
┌─────────────────────────────────────────────────────────────────┐
│  gecko-trade-agent  →  local advisor (your keys, your laptop)   │
│  (cache-then-charge; journal in Mongo; advisor mode in v0.1)    │
└─────────────────────────────────────────────────────────────────┘
                              ↓
┌─────────────────────────────────────────────────────────────────┐
│  Execution adapters (v0.2): okx · sendai · backpack             │
│  (venue-neutral; agent doesn't pick the wallet provider)        │
└─────────────────────────────────────────────────────────────────┘

The agent runtime never reads the corpus directly — it calls the oracle. The oracle never touches keys. The execution layer never picks a verdict. Each layer is independently swappable.


Sub-agents (.claude/agents/)

After a verdict lands, five pre-baked personas help you act on it:

Agent Role
research-analyst Explore the verdict; run free gecko_ask follow-ups.
market-validator Adversarial reader of the verdict + dissent.
technical-architect Translate strategy spec into runnable code.
validator Pre-execution sanity check.
builder Scaffolds the integration.

Cost transparency

Per gecko_trade_research basic at $0.25 retail:

Line Real cost
LLM (5-voice debate, gpt-4o-mini) $0.012
Embeddings (text-embedding-3-small) $0.002
Retrieval (Mongo Atlas Search) $0.001
Total $0.015
Margin $0.235 (94%)

Run gecko-mcp economics <session_id> after any session for the full breakdown.


Examples


Roadmap

  • v0.1 (now) — trade oracle live, advisor mode, frames.ag settlement, investor-canon corpus.
  • v0.2 — trader mode (signed execution via okx / sendai / backpack adapters); licensed canon (O'Reilly, Perlego) behind a tier flag.
  • 🔜 v1.0 — multi-agent daemon mode; creator-attribution settlements on-chain to canonical-source contributors.

Sister repos


License

MIT. See LICENSE.


Strategy oracle for autonomous trading agents · geckovision.tech · No API keys, just a wallet.

About

Repository for the Gecko claude skills - It has all the necessary config to setup your project

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors