solana-token-safety-skill

A second opinion on whether a Solana token's market is real — before an agent trades it or a builder lists it.

Built for the Solana AI Kit. A progressive, token-efficient skill that gives any coding/trading agent a one-call pre-trade safety gate.

Try it in 30 seconds

No key, no install — hit the live oracle:

curl -s -X POST https://api.geckovision.tech/safety \
  -H "Content-Type: application/json" \
  -d '{"mint":"JUPyiwrYJFskUPiHa7hkeR8VUtAeFoSYbKedZNsDvCN"}'
# -> {"gate":"ok", ... "information_mev":{"label":"clean", ...}}

…or run the zero-dependency reference client (prints the verdict, exits non-zero on a block so it composes in CI):

python scripts/safety.py JUPyiwrYJFskUPiHa7hkeR8VUtAeFoSYbKedZNsDvCN
# ✅  gate = OK  (JUPyiwr…)
#     - No manipulation signals: liquidity 0.40% of market cap.

The problem it solves

Every contract checker on Solana (RugCheck, GoPlus, Solsniffer, Bubblemaps) tells you whether the contract is safe — mint/freeze authority, honeypot bytecode. They routinely pass tokens whose price and volume are manufactured: bot- inflated wash volume, a fake market cap held up by a few thousand dollars of liquidity, the same token quoted far above its real price in thin satellite pools, one wallet sitting on the float.

That's the gap that drains agents and rugs retail. A static contract scan can't see it, because the manipulation is in the market data, not the contract.

This skill is that missing layer. It flagged a live token (BrCA) as block / manipulated 0.95 — $5.73M quoted market cap on $6.5K of liquidity, top holder 72% — while the platform it traded on rated it "Normal".

What it does

One call → a one-glance gate:

block    — do not trade / do not list (fake mcap, manipulated/wash, depeg, honeypot)
caution  — reduced size + explicit confirm (thin liquidity, concentration, elevated)
ok       — checked and clean
unknown  — the check could not run (fail-OPEN — never assume safe)

…plus the reasons[] to explain the verdict to a human, and (as the Launch- Firewall stream rolls out) a wash_risk block with the specific flow signal that fired (one-sided buy-loop, multi-pool price bait, wash recirculation, sybil funder cluster).

It's free, keyless, and fail-OPEN — POST a mint, get a gate.

Install

git clone <this-repo> && cd solana-token-safety-skill
bash install.sh          # copies the skill + /safety-check command into ~/.claude

Or point it at a custom config dir: CLAUDE_DIR=/path bash install.sh.

Use

• Ask your agent: "is this token safe to trade: <mint>?"
• Trader command: /safety-check <mint> · Issuer command: /launch-check <mint>
• Agent persona: token-safety-auditor runs the full diligence flow
• Wire it into an agent's pre-trade loop → skill/agent-integration.md

Two sides, one check

• Agents — gate a buy/swap before you sign (/safety-check, the pre-trade-safety rule).
• Issuers / launchpads — vet a token's launch for bot manipulation in the first candles (/launch-check, the Block-Zero read).

Structure

skill/
  SKILL.md                 entry point + routing (load this)
  quickstart.md            the one call (curl / Python / TS)
  signals.md               what each flag/signal means
  interpreting-verdicts.md gate semantics + fail-OPEN + FP awareness
  agent-integration.md     MCP + HTTP pre-trade gate pattern
  launch-check.md          issuer / Block-Zero read
commands/
  safety-check.md          /safety-check <mint>   (trader)
  launch-check.md          /launch-check <mint>   (issuer)
agents/
  token-safety-auditor.md  market-integrity specialist persona
rules/
  pre-trade-safety.md      opt-in: gate token swaps with a check first
examples/
  clean-jup.json           real OK verdict (verified) + blocked-brca.json (illustrative)
scripts/
  safety.py                zero-dependency reference client (stdlib urllib)
  safety.ts                zero-dependency reference client (built-in fetch)
install.sh                 copy-only installer (no binaries, no network)
CLAUDE.md · LICENSE (MIT) · README.md

How it works

The skill is a thin, safe client over Gecko's free /safety oracle (POST https://api.geckovision.tech/safety) or the gecko_safety MCP tool. The oracle reads on-chain liquidity/holder/market data, computes deterministic manipulation signals, and returns the gate. The skill ships no executable logic of its own beyond the install copy — nothing opaque, nothing to audit beyond markdown.

• Docs: https://docs.geckovision.tech
• Provider: Gecko — a decision-integrity oracle for Solana agents. The /safety tier is free by design; the considered multi-voice adjudication lives behind the paid /trade_research tier (not used by this skill).

License

MIT — ready to be merged or submoduled into the kit.