Everyone built hands, eyes, and ears. We built the brain
Ocean Agent is an AI native trading layer for Pacifica. Tell it what to trade; it fixes the tick, lot, and min order, then executes correctly.
It doesn't debate. It doesn't guess. It measures, then studies its own results.
Community & Support
Discord
Join the community. Get help, share progress, and build on Pacifica.
Twitter / X
Follow @pacifica_fi on X for the latest news and updates.
Office Hours
KST 09:00 to 18:00
UTC 00:00 to 09:00
US Eastern (ET) 20:00 to 05:00
What Ocean Agent is
An AI native trading layer for Pacifica. It comes in two parts: an MCP server your AI client calls on demand, and a 24/7 autonomous entity that trades on its own inside a policy you set.
The official Pacifica MCP passes the API through as is: your AI has to compute exact prices and sizes, and a wrong tick or lot gets rejected. Ocean Agent adds the layer above that: tick, lot, and min order correction, risk based sizing, a confirm gate, measured statistics, and self learning.
| At a glance | |
|---|---|
| Exchange | Pacifica (Solana perpetual futures) |
| Interface | MCP server + standalone 24/7 autonomous process |
| Signing | Ed25519 API key, trading only, cannot withdraw funds |
| Markets | Every Pacifica market: crypto, stocks, commodities, FX |
| Decision basis | Measured edge (backtest matrix + live win rates), not a model's guess |
| Status | Live on mainnet with small size since July 2026, building the public track record |
From zero to running
Live on PyPI. Nothing to preinstall, one command sets everything up. The autonomous bot is optional.
| Step | What you do |
|---|---|
| 1 · API key | Create a Pacifica agent key at app.pacifica.fi/apikey. It can trade but cannot withdraw, and you can revoke it any time. |
| 2 · One command | Paste the installer command, answer two questions. It sets up everything, including Python and Claude's config. |
| 3 · Chat | Restart Claude and the trading tools appear, analyze charts, check positions, trade by talking. No bot required. |
| 4 · Run the bot (optional) | Only if you want autonomous trading: start the entity and it reads the market, enters only setups that clear every gate, and manages stops around the clock. |
Start in dry mode until you understand exactly what it does.
Install
One command. Paste it into a terminal, answer two questions (wallet public address, API key), it installs everything and connects Claude automatically.
powershell -ExecutionPolicy ByPass -c "irm https://oceanagent.vercel.app/install.ps1 | iex"sh -c "$(curl -LsSf https://oceanagent.vercel.app/install.sh)"It asks for your wallet public address and an agent key from app.pacifica.fi/apikey (trade-only, cannot withdraw, revocable any time). When it finishes, restart Claude, the trading tools appear in chat. No bot required.
Works in your language. You chat with your AI in any language: Korean, Japanese, Spanish, anything. It drives the tools and explains every result in your language, and new-version notices arrive in it too.
uv run --with ocean-agent python -m ocean_agent.autonomous --dryOnly for hands-off trading. --dry decides but places no orders; drop it when you trust it.
Manual setup (any MCP client)
.env file with your keysADDRESS=your_wallet_address
PACIFICA_API_KEY=your_agent_key
PACIFICA_BASE_URL=https://api.pacifica.fi{ "mcpServers": { "ocean-agent": {
"command": "uvx", "args": ["ocean-agent@latest"],
"env": { "PACIFICA_ENV_FILE": "/absolute/path/to/.env" } } } }@latest means every client restart checks PyPI and picks up new releases automatically. Your local data and keys are never touched by updates.
uv run --with ocean-agent python -m ocean_agent.doctorHow Ocean Agent compares
Seven AI trading agents in one table. Every other one either lets a model guess or debate, or piles on rules. Here is what actually separates them, unpacked.
How it decides the one that makes money
Most agents let a language model read the chart and form an opinion, or run several personas that debate. Ocean Agent trades on a measured edge: for each coin, timeframe, and signal it holds an expected value net of fees, from a matrix of past measurements. If that number is not positive, it does not trade. A guess can be confident and wrong; a measured edge is just a number that either clears the bar or does not.
Self learning unique here
Backtests go stale as the market changes character. On a schedule, Ocean Agent re measures the full matrix of coins, timeframes, and signals, and updates which ones it trusts. In one measurement an 8 hour band showed no edge at all; weeks later it was the best. No competitor re measures its own edge, so their parameters quietly rot.
Meta learning unique here
It tracks how often its own predicted win rate actually held. When it was too optimistic, it corrects the next estimate down. In other words it learns how wrong it tends to be and fixes it. Nothing else in this set does this.
Win / loss attribution unique here
Every closed trade is graded: did the signal work, or did the market just drift your way? A lucky win is not counted as a good signal, and a good signal that lost to a broad move is not thrown away. This is how it avoids learning the wrong lesson from noise.
Market & custody
Leverage perpetuals on crypto, stocks, commodities, and FX, all on Pacifica. It is non custodial: a trading only API key that cannot withdraw your funds. Several competitors are spot only, stocks only, or custodial.
Autonomy & control
It runs 24/7 as its own process, so it holds positions and manages stops even with the AI offline. Take profit and stop loss are exchange native and trail automatically. Alerts go to Telegram, with two way control on the roadmap.
Where we are honestly behind
Setup is a config file and a key, not a single click. There is no regulation or insurance wrapper. The live mainnet track record is still young, it trades small until the numbers justify more. And it is a solo open source project, not a funded company. The table below shows those gaps in the same view as the wins.
| Ocean AgentPacifica | Coinbasefor Agents | RobinhoodAgents | Minaramulti chain | TradingAgents | EVAProtocol | Virtualson chain | |
|---|---|---|---|---|---|---|---|
| Decision quality | |||||||
| How it decides | Measured stats | Model guess | Model guess | User rules | Model debate | Model debate | Templates |
| Self learning | re measures | ✕ | ✕ | ✕ | reflection | ✕ | ✕ |
| Meta learning | unique | ✕ | ✕ | ✕ | ✕ | ✕ | ✕ |
| Win/loss attribution | unique | ✕ | ✕ | ✕ | ✕ | ✕ | ✕ |
| Market and custody | |||||||
| Instrument | Leverage perps | Spot | Spot stocks | Perps and spot | Stocks | Multi | Perps |
| Stocks and commodities | yes | soon | stocks only | ✕ | stocks only | ✕ | ✕ |
| Non custodial | yes | ✕ | ✕ | yes | optional | yes | yes |
| Autonomy and control | |||||||
| Runs 24/7 with AI offline | own process | ✕ | ✕ | rules | yes | ✕ | yes |
| Native TP/SL and trailing | auto | manual | manual | manual | ✕ | ✕ | yes |
| Telegram two way control | alerts now | ✕ | ✕ | ✕ | ✕ | ✕ | yes |
| Where we are behind | |||||||
| Ease of setup | config + key | 1 login | 1 command | simple | docker | self host | self host |
| Regulation and insurance | none | KYC | SIPC | ✕ | ✕ | ✕ | ✕ |
| Live track record | live · small | live | live | live | ✕ | ✕ | ✕ |
| Brand and capital | solo OSS | public co | public co | funded | OSS | token | token |
Measured edge, not opinion
A signal only trades if it beats the market's own drift, on that specific coin and timeframe, net of fees. Everything a decision touches is a measured number.
| Input | What it contributes |
|---|---|
| Baseline subtracted edge | The signal must beat unconditional market drift, not just win often. |
| Look ahead prevention | The unclosed candle is dropped, so no signal reads a bar still forming. |
| Backtest matrix | Expected value per coin, timeframe, and signal, net of fees. |
| Live win rate | Observed win rate from the agent's own trades, weighted higher as it grows. |
| Regime read | Fear and Greed context; flags signals whose edge flips in the current regime. |
| Conviction | One score that fuses the above and sets entry order and size. |
It studies its own results
The loop that makes it better over time. It learns only from its own signal trades, so nothing else pollutes the measured edge.
| Layer | What it does |
|---|---|
| Re measurement | On a schedule it re measures the full matrix of coins, timeframes, and signals, and updates which it trusts. Regimes change, so fixed parameters go stale. |
| Meta learning | Tracks how often its predicted win rate actually held, and corrects its own overconfidence once enough samples exist. |
| Post mortem | Grades each closed trade as skill or luck, so a lucky win is not mistaken for a good signal. |
| Adaptation | Bans signals that lose in live grading, trims size in drawdown, restores it on recovery, always within policy bounds. |
Runs while you sleep, never runs blind
A separate always on process governed by a policy file, a delegation contract it cannot act outside of.
Capital split across trading, carry, and a cash reserve, rebalanced each cycle.
Enters only setups that pass its win rate gates. Every position gets native TP and SL.
Rotates through every Pacifica market, building a win rate database at zero risk.
Moves the stop to breakeven when ahead, trails it as profit grows, takes partial profit at target.
Pauses new entries on a bad day, and flattens fully on a deep drawdown.
It acts only inside the policy file. Nothing outside those bounds.
Thirteen things the passthrough doesn't have
Each one answers a question a trader actually asks, with data instead of an opinion.
Funding scan
Ranks every market by funding APR.
One command carry
Spot buy and perp short in one atomic batch.
Carry unwind
Closes both legs together. No half open exposure.
OI hedge design
Sizes a cross exchange hedge with fees and funding included.
Bracket entry
Entry plus exchange native TP and SL in one call.
Protect position
Adds native TP and SL to any open position.
Chart analysis
Four timeframes with a measured hit rate per signal.
Top setups
Live ranking by expected value, win rate, and sample size.
Market regime
Reads Fear and Greed, blocks bad trades for the regime.
Print verdict
Tells you whether a Print offer is worth taking.
Learned win rates
A win rate database grown from live observation.
Prediction grading
Past calls against what actually happened.
Safety layer
Confirm gate, automatic price correction, separate testnet and mainnet keys, and trading only keys that cannot withdraw.
The plays it can run
Every strategy is measured and gated. If it can't beat the existing logic, it doesn't ship.
| Strategy | What it does · when it fits |
|---|---|
| Statistical setup | The main engine. Enters signals that clear a measured win rate and expected value gate for the current regime. |
| Bracket entry | Entry plus exchange native take profit and stop loss in one call. The stops live on the exchange and fire even with your machine off. |
| Funding carry | Delta neutral: spot buy plus perp short, filled atomically. Best on larger capital; small accounts skip it. |
| Regime filter | Flags setups whose historical edge flips sign in the current market regime. |
| OI hedge | Sizes an open interest position with its cross exchange hedge, fees and funding included. |
One policy file governs everything
The autonomous entity acts only inside these bounds. Widen them carefully.
| Setting | Meaning |
|---|---|
| Capital | Fixed amount, or live to size to your actual wallet balance. |
| Risk per trade | How much of capital a single stop may cost. |
| Max leverage | A ceiling the policy enforces over the market maximum. |
| Max concurrent | How many positions may be open at once. |
| Soft / final stop | Pause new entries on a bad day; full flatten on a deep drawdown. |
| Allocations | Split between trading, carry, and cash reserve. |
| Mode | Aggressive (harvest data) or careful (proven setups only). A one line switch. |
Built so mistakes are structurally hard
The safety layer sits under every money moving action.
| Guard | What it prevents |
|---|---|
| Two step confirm | Every money moving tool previews first and executes only on explicit confirmation. |
| Trading only keys | API keys can trade but cannot move funds out. Revocable any time. |
| Key separation | Testnet and mainnet keys and data are kept apart. |
| Hard sizing guard | Any order above a notional ceiling is rejected at the API layer, whatever the source. |
| Liquidity gate | Skips markets where your own order would be a large share of daily volume. |
| Native stops | Exchange native TP and SL on every entry, so they fire with your machine off. |
Risk. This is trading software. Leveraged perpetuals can lose more than the margin you post. Measured win rates come from historical data and are regime dependent. Nothing here is financial advice. Only risk what you can afford to lose.
Capability grows only as live evidence earns it
Correctness and risk first. Each phase unlocks only when a live data gate is met.
Where we win, and where we honestly don't
Where we win
- A statistics brain that measures its edge
- Grades its own trades: skill or luck
- Corrects its own overconfidence
- Live leverage perps on stocks and commodities
Where we are behind
- Setup is config and a key, not one click
- No regulation or insurance
- Live track record is still young (mainnet since July 2026)
- A solo open source project
Common questions
No. It uses a trading only API key with no withdrawal rights, and you can revoke it any time.
Yes: live on mainnet with small size since July 2026. Sizes stay small until the live track record justifies more.
Yes. You chat with your AI in any language; it drives the tools and explains the results in that language. Raw tool reports and this site are in English, but your AI translates as it goes.
The autonomous entity needs its own process running, but stops are exchange native and fire even if your machine goes offline.
Every Pacifica market: crypto, stocks, commodities, and FX, as leverage perpetuals.
Not until the live track record is real. That is the one number it will never fake.
What changed, and what we measured
Every entry states the change and the measurement behind it. When a measurement said the current setting was already better, we kept it and wrote that down too. Backtest figures assume perfect fills and exclude funding. They describe past regimes, not forecasts.
Routine security review
Version 0.3.6. We ran a full security review against the current landscape of crypto and supply-chain attacks, and tightened a few things as a result. No issue was ever exploited.
The design principles that keep funds safe were confirmed intact: the trading key is trade-only and can never withdraw, every order that moves money passes an explicit confirmation and a hard size ceiling, and the agent only ever signs requests it built itself. Your keys stay on your machine and are never transmitted to us.
We review security on an ongoing basis and fold any hardening into the next release. The best thing you can do is stay current: existing installs pick up the latest version automatically on the next client restart.
Wider stops, measured; and a consent-based developer fee
Version 0.3.4. The stops were sitting inside the noise band and cutting winning trades, so they are now twice as wide. Installs also ask, once, about a small developer fee. Nothing is approved silently.
What changed
| Change | Measured basis |
|---|---|
| Stop width doubled | 24h-horizon picks swing about 3% against entry before going the right way, and the old stops sat inside that noise. Sweeping multipliers on a 7-year replay: 1.0x $8,048, 1.5x $9,134, 2.0x $9,709, 2.5x $8,894, 3.0x $8,606. A single peak at 2.0, not a runaway curve. Position size auto-halves because sizing is stop-based, so risk per trade is unchanged, and max drawdown fell from 5.9% to 5.2%. |
| Harness parity bug fixed | The backtest harness still carried a fixed swing stop block that the live bot withdrew on Aug 5. It was removed before the sweep above, so measurements now replay exactly what the bot runs. The default harness reproduces the live config to the cent. |
| Terms acceptance at install | The installer asks once, in your OS language, whether you have agreed to the Terms of Use. Accepting continues the install and signs the builder-fee approval the Terms describe (0.01%, 1 bps, section 11); declining cancels the installation, the standard EULA behavior. AI-client-only users get the same one-time question in chat, in the conversation language. Nothing is ever approved without your explicit answer, and an approval can be revoked anytime in your Pacifica account. |
| Install page terms gate | The install commands unlock after the full scrollable Terms of Use (18+, unregulated software, total-loss risk, fees). Agreeing also offers the commands to copy right in the window. |
Tried and rejected, on the numbers
Four plausible ideas were measured this round and kept out. Fixed take-profit exits cut the 30-day pick total from +16.6% to +5.4% by amputating the big winners. Break-even floors after a partial move did worse than holding. Dropping in-progress bars changed accuracy by 0.1 points, noise. And the popular idea that the 4-hour bar should veto smaller ones inverted on our data: picks that fought the 4h bar hit 55.2% against 49.4% for picks that agreed with it. Each rejection is recorded so we do not walk into the same temptation twice.
Existing installs pick this up on the next client restart. The bot applies the new stops from its next start.
Ocean Agent is live on PyPI
The agent is now a public package. One command installs it, registers it with Claude, and keeps it updated.
Until now the only way to run it was to clone the repository. It now ships on PyPI as ocean-agent, currently at version 0.3.3. Install with the one-line installer on this site, which sets up uv, Python, your keys and Claude's MCP config in a single paste, or with pip install ocean-agent if you already have a Python environment.
What is in this release
| Change | Measured basis |
|---|---|
| Renamed: mustache-mcp is now ocean-agent | Existing user data migrates automatically on first run. |
| Win-rate gate 52% → 45% | Nine years of walk-forward measurement put the composite-vote ceiling at 54.3%, so the 52% per-signal gate was starving entries: two days, zero trades. A 7-year replay showed +21% account improvement at 45%. The pure-edge +2%p gate stays, so the floor holds. |
| Volume and liquidity filters removed, scan widened to all 100 markets | Measurement showed the filters were not what was blocking entries. |
| Regime filter off by default | 30-day A/B: +$6,052 with it off against +$2,612 with it on, and lower drawdown. |
| Margin checked before every entry | Fail-closed: if the check cannot run, the entry does not go out. |
| Defense line ratchets with equity peaks | The stop-trading floor now follows new highs instead of sitting at its starting value. |
| Seed data bundled | The 9-year matrix and the calibration curve ship inside the package, so a new install starts with measured gates rather than empty ones. |
Reliability and plumbing
Seven fund-moving tools are now marked destructive in the MCP layer, so your client asks before they run. Errors are standardized, rate limits back off on 429, and remeasured timeframes hot-reload into a running session instead of waiting for a restart. A timezone bug in prediction grading is fixed, expired predictions stop retrying after a cap, orphan positions raise an alert, and a project-wide lint sweep removed 13 dead-code findings.
Updates arrive on their own
Installs register the server as ocean-agent@latest, so every client restart picks up the newest release. A session that is already running shows a one-time in-chat notice when a newer version appears on PyPI.
Full backtest harness, and seven bugs it found
We replayed the entire trading pipeline over eight years of price history and let the agent grade itself.
Until now the agent could measure whether a signal predicted direction, and what a signal's average edge was. Neither answered the question that matters: what happens to the account. Stops, trailing, concurrency and budget change the answer completely.
Roughly half the trades, more total profit. Fewer fees, better entries.
Bugs found and fixed
| What was wrong | Why it mattered |
|---|---|
| Stops judged on closing price only | A bar that pierced the stop and recovered was scored a win. Stop-out rate read 55% when it was really 68%. |
| Timeframes ranked by average edge | The agent only trades its best signals, so ranking by the average picked nearly the reverse order. |
| Gate used a different stop than the trade | Setups cleared the expected-value gate on one stop distance, then traded with a wider one. |
| Swing stop applied to every timeframe | A 1.5% stop tuned on hourly bars was applied to daily bars, where a day moves 3–5%. |
| Risk setting never reached sizing | The policy file could be edited freely and nothing changed. |
| Scan universe hardcoded to 12 | 29 markets passed the liquidity floor; 17 were never looked at. |
| Re-measurement never reached the agent | Timeframe selection was computed once at startup and never refreshed. |
Settings changed
| Setting | Before → after | Measured basis |
|---|---|---|
| Path judgment | close → high / low | Stop-out rate 55% → 68% once intrabar wicks count |
| Timeframe selection | average → top 3 | Daily ranked last by average, first by top-3 |
| Timeframes traded | 5m/15m/1h → 8h/12h/1d | Hourly lost money per trade |
| Partial take-profit | on → off | More profit and less drawdown (5.8% → 4.5%) |
| Binance history | 12 → 18 markets | Sample counts grew 4–9× |
| Position cap | $150 → $250 | +40% total profit |
| Minimum margin | none → $25 | Skips positions too small to be worth the round trip |
| Scan universe | 12 → 40 markets | The liquidity floor already filters |
Tested and deliberately not changed
Nine ideas measured worse than what was already there. Recording them matters as much as the changes — it stops us re-litigating settled questions.
| Idea | Result |
|---|---|
| Fixed 1.5% stop / 5% target | Negative even under the most generous fill assumption |
| Volume-spike entry filter | Entering on volume spikes was worse by 1.8pp |
| Lower the sample requirement | 30 samples beat 20, 15, 10 and 5 |
| Cap holding time | A 6-hour cap erased 85% of profit |
| Size as a % of equity | Compounding blow-up — impressive on paper, impossible in practice |
| Tighten the entry gates | Worse at three consecutive thresholds |
| Shrink targets to correct bias | Narrower targets removed the room trailing stops need |
What actually produces the returns
| Prediction | Accuracy | Role |
|---|---|---|
| Direction — up or down | 50.2% | Essentially none. Gone entirely one week out. |
| Magnitude — how far | 0.265 rank corr. | This is the edge. It places the target and the stop. |
Direction is close to a coin flip. Knowing how far price tends to travel is what lets the agent place an asymmetric target and stop — and the trailing stop converts that asymmetry into profit. Disable trailing and 92% of the returns disappear, while drawdown triples.
It is not a forecaster. It is a trend follower with a measured sense of scale.
MCP server audit, and Print closed out
The MCP server had been pointed at testnet. Print was measured across every structure and set aside.
The MCP server was pointed at testnet
Tools reported a play-money balance while the agent traded real funds, and logged predictions into the wrong network's files. One line in the config; the consequences were not small.
MCP tools now share the agent's policy
The setup scanner was sizing with the exchange's maximum leverage instead of the policy's, so its margin figures were an order of magnitude away from what the agent would actually do.
Print evaluated and set aside
Pacifica's Print product pays a premium while a target order waits. We measured every structure we could think of — single side, both sides, perp-hedged, distance sweeps, leverage sweeps, BTC+ETH mixes, signal timing. All negative.
The reason is one line: implied volatility (26–38%) sits below realised volatility (40–53%). Selling that is a losing trade regardless of structure. One condition flips it — when trailing 7-day volatility drops below 24.2%, ETH shorts at 2% distance measured +17%/yr. That window is open about 3% of the time, so the agent watches for it instead of trading it.
Separating what it knows from what it hopes
Trade results were attributed to the wrong trade, and risk sizing was running 12× smaller than intended.
Profit and loss attributed to the wrong trade
A symbol's lifetime P&L was being read as the result of the most recent trade. One winning BTC position was recorded as a $7.30 loss. Re-grading the history turned 1 win / 7 trades into 4 wins / 7 trades — which finally matched the account.
Risk sizing was 12× off
The scanner computed margin using the exchange's leverage cap while the executor placed orders at the policy cap. Intended risk of 2% per trade was operating at 0.17%.
Auto-bans are no longer permanent
Five samples and a 45% floor was killing genuinely-52% signals 46% of the time, with no way back. Bans now expire.
Terms of use
Last updated: August 7, 2026 · Effective immediately for all users. Please read these terms carefully before downloading, installing, or using the Ocean Agent software or this site. By doing any of those things you agree to be bound by these terms. If you do not agree, do not use the software.
1 · Introduction and definitions
These Terms of Use ("Terms") govern your use of the Ocean Agent software and of this website. In these Terms:
- "Software" means the ocean-agent package published on PyPI, including its MCP server, its autonomous trading process, its command-line utilities, its bundled measurement data, and this website's install scripts;
- "Authors" means the developers and contributors of the Software;
- "Exchange" means Pacifica, the third-party trading venue the Software connects to;
- "Agent Key" means the trade-only API key you create at the Exchange and supply to the Software;
- "AI Client" means any MCP-compatible application (such as Claude) that you connect to the Software;
- "Autonomous Mode" means the optional always-on process that scans markets and places orders without per-order confirmation;
- "You" means the person installing or using the Software.
The Software is open-source software released under the MIT License. It is a tool that runs entirely on your own computer. The Authors do not operate a service, do not run servers that touch your funds or orders, and have no ongoing relationship with you beyond publishing the code. The Software is not affiliated with, endorsed by, or operated by the Exchange.
2 · Unregulated software: no protections
No financial regulator in any jurisdiction has licensed, reviewed, or approved the Software or the Authors. The Authors are not a bank, exchange, broker-dealer, investment adviser, fund manager, money transmitter, or wallet provider. Using the Software gives you none of the protections that apply to regulated financial services: no deposit insurance, no investor compensation scheme, no segregation of client assets, and no financial ombudsman or complaints tribunal you can escalate to. If something goes wrong, the realistic remedies are limited to those in Section 15.
3 · Eligibility
You may use the Software only if all of the following are true:
- you are at least 18 years old and legally able to enter into binding contracts;
- you are not located in, and do not access the Exchange from, a jurisdiction where trading leveraged cryptocurrency derivatives is restricted or prohibited, including any jurisdiction the Exchange itself blocks or restricts (the Exchange's own list, which may change at any time, controls);
- you are not a resident or citizen of, and are not accessing from, any country subject to comprehensive sanctions, and you are not on any sanctions or restricted-party list (including OFAC, UN, and EU lists);
- you will not use VPNs or other means to disguise your location in order to access the Exchange from a restricted jurisdiction;
- your use of the Software and of the Exchange is lawful where you live.
Determining all of the above is solely your responsibility. The Software performs no identity verification and cannot check any of it for you.
4 · No accounts, no registration, no data collection
The Software has no sign-up, no login, and no user database. The Authors do not collect, store, or transmit your personal data, your keys, your balances, or your trading history, everything the Software records lives in files on your own computer. This site is a static page and sets no tracking of its own. The only parties that see your activity are the Exchange (which you have a direct relationship with under its own terms) and, if you use an AI Client, that client's provider.
5 · Non-custodial design: wallets and keys
The Software never holds your funds and never sees your wallet's private key or seed phrase. It signs orders locally with an Agent Key that is trade-only by design: it cannot withdraw funds, and you can revoke it at the Exchange at any time. Signed requests go directly from your computer to the Exchange, there is no intermediate server.
Because of this design, the Authors cannot: access your account, freeze or unfreeze anything, recover keys you lose, cancel or reverse orders, or return funds. There is no support desk that can undo a trade. Anyone who obtains your Agent Key can trade your account until you revoke it; keeping it secret, and keeping the computer it lives on secure, is entirely your responsibility.
6 · The MCP server and AI clients
The Software's MCP server exposes trading and analysis tools to the AI Client you connect. You acknowledge that:
- AI language models can misunderstand instructions, make mistakes, and state wrong things confidently. Review what your AI Client is about to do before confirming orders;
- tools that move money require an explicit confirmation parameter, but your AI Client controls how that confirmation is presented to you;
- the quality of decisions made through an AI Client depends on that client and its model, neither is under the Authors' control;
- analysis tools output measurements of past data. They are not predictions and not advice (Section 14).
7 · Autonomous Mode
If you start Autonomous Mode with a trading key, the Software will place real, leveraged orders by itself, around the clock, without asking you first, that is its purpose. It operates inside a policy file you control (capital, risk per trade, leverage cap, stop rules), but within those bounds its decisions are its own. You are responsible for every order placed under your key, including orders you never personally reviewed. Run it in dry mode (decisions logged, no orders) until you understand exactly what it does, keep the policy file conservative, and check on it regularly. The Software adapts parameters within policy bounds; it never rewrites its own code.
8 · Experimental features and undocumented endpoints
Some features, notably Print, rely on Exchange endpoints that are not officially documented. They may change or stop working at any time without notice, and their behavior in edge cases is not guaranteed by anyone. Features marked experimental are provided for evaluation only; using them with real funds is at your own heightened risk.
9 · Orders are irreversible: no refunds
Orders executed on the Exchange are final. Cryptocurrency transactions cannot be recalled, and leveraged positions can be liquidated before you have a chance to react. Because the Software never holds funds, the Authors have nothing to refund from and cannot issue refunds, reimbursements, or compensation under any circumstances, including losses caused by bugs, misconfiguration, compromised keys, or fraud by third parties.
10 · Access is controlled by the Exchange, not the Authors
The Authors have no power to suspend or terminate your trading, and equally no power to restore it. The Exchange may restrict, suspend, or close your account under its own terms at any time, may delist markets, change fees, or halt trading. None of that is within the Authors' control, and the Software may stop working, in whole or in part, whenever the Exchange changes its API.
11 · Fees
The Software itself is free to install and use. All trading fees, funding payments, and spreads are set and collected by the Exchange under its own schedule, which can change. The Software's expected-value calculations account for fees as measured at the time; actual fees may differ.
Developer fee (builder code). During installation you are asked whether to approve a builder fee of 0.01% (1 bps) of each filled order's notional value, paid to the Authors through the Exchange's builder-code program. Approval is optional and takes effect only with your own cryptographic signature; the Software never approves it silently. Declining does not change any functionality. You can revoke an approval at any time in your Exchange account settings, and the Exchange rejects any order whose builder fee would exceed the cap you signed.
12 · Taxes
Trading profits and losses may create tax obligations where you live. Determining, reporting, and paying any such taxes is solely your responsibility. The Software's records are kept for its own learning loop and are not tax documents; do not rely on them as complete or authoritative for tax purposes.
13 · Privacy
Everything the Software learns (trade history, win-rate databases, measurement matrices, logs) is stored locally on your machine and never transmitted to the Authors. There is no telemetry, no analytics, and no phone-home. Your network traffic goes to the Exchange (to trade) and to public package registries (to install). If you optionally configure notifications, those go to the provider you configure.
14 · Disclaimers and risk statement
No advice
Nothing the Software or this site outputs is investment advice, a recommendation, an offer, or a solicitation to trade. Win rates, expected values, backtests, forecasts, and "recommended settings" are measurements of past price data produced by code, not predictions, and not tailored to you. The Authors are not licensed financial advisors and know nothing about your circumstances. If you need advice, consult a licensed professional.
Market risk
Cryptocurrency prices are extremely volatile and trade around the clock. Leveraged perpetual futures amplify both gains and losses: positions can lose more than the margin posted and can be liquidated entirely. Stop-loss orders are not guarantees, in a fast market or a price gap the fill can be far worse than the stop price, or the stop may not fill at all. Thin order books add slippage that can exceed the edge being traded.
Strategy risk
The Software's measured edges are small, measurement puts the realistic directional ceiling in this market near 54~58%, and backtested figures above that did not survive out-of-sample testing. Edges that small are fragile: they can vanish when the market regime changes and are easily erased by fees and slippage. Past measurement does not guarantee future results. A period of profits is not proof the system works; a period of losses is possible at any time.
Technical risk
Software has bugs. The Software may miscalculate, crash, lose connectivity, or behave unexpectedly after an Exchange API change. Your own machine may fail, sleep, lose power, or be compromised. The Exchange may go down while you hold positions. Exchange-native stops remain on the Exchange's servers if your machine goes offline, but everything else the Software manages stops with it.
As is
The Software and this site are provided "as is" and "as available", without warranty of any kind, express or implied, including merchantability, fitness for a particular purpose, accuracy, non-infringement, or uninterrupted operation.
15 · Limitation of liability
To the maximum extent permitted by applicable law, the Authors and contributors shall not be liable for any loss or damage of any kind arising out of or in connection with the Software or this site, including, without limitation, trading losses, lost profits, liquidations, losses caused by software defects, incorrect calculations, stale data, Exchange outages or API changes, network failures, misconfiguration, compromised credentials, actions taken by an AI Client, or security incidents on your machine, whether in contract, tort, or otherwise, even if advised of the possibility of such damage. Where liability cannot be excluded by law, it is limited to the amount you paid for the Software, which is zero. Nothing in these Terms excludes liability that cannot lawfully be excluded, such as liability for fraud.
16 · Your representations and warranties
By using the Software you represent and warrant that:
- you meet the eligibility conditions in Section 3;
- the funds you trade were acquired lawfully;
- you have the technical ability to install, configure, and supervise the Software, or accept the consequences of not supervising it;
- you understand leveraged derivatives and the risks in Section 14;
- you are acting for yourself, not on behalf of another person or entity;
- you will not use the Software for market manipulation, money laundering, sanctions evasion, or any unlawful purpose.
17 · Term and stopping use
These Terms apply from your first use of the Software until you stop using it. You can stop at any time: close positions, revoke your Agent Key at the Exchange, and uninstall. Revoking the key immediately removes the Software's ability to trade your account. Sections that by their nature should survive (disclaimers, limitation of liability, and governing terms) survive termination.
18 · Security obligations
You are responsible for: keeping your Agent Key and .env file secret and out of version control and chat messages; securing the computer the Software runs on; reviewing the policy file before running Autonomous Mode with real funds; revoking keys promptly if you suspect compromise; and monitoring positions the Software opens. If your key leaks, revoke it at the Exchange immediately, that invalidates it at once.
19 · Force majeure
The Authors are not responsible for any failure or delay caused by events beyond reasonable control, including natural disasters, war, civil unrest, labor disputes, power or internet outages, actions of governments or regulators, Exchange failures, blockchain congestion or forks, or failures of third-party infrastructure.
20 · Changes to these Terms and to the Software
These Terms may be updated as the Software evolves, to reflect new features, removed features, legal requirements, or clarity improvements. The current version always lives on this page with the date above; material changes will be reflected in the project changelog. Continuing to use the Software after a change means you accept the updated Terms. The Software itself may change at any time: features can be added, altered, or removed, and published measurement data can be revised as new measurements replace old ones.
21 · Open-source license and intellectual property
The Software's source code is licensed under the MIT License, which governs your rights to use, copy, modify, and redistribute the code. These Terms do not restrict rights the MIT License grants you; they govern your use of the Software against a live exchange account and your use of this site. Trademarks, names, and logos of the Exchange and of AI Client providers belong to their respective owners. You are encouraged to read the source before connecting real funds, it is public precisely so you can verify what it does.
22 · General
Entire agreement. These Terms and the MIT License are the entire agreement between you and the Authors concerning the Software, replacing any prior statements. Severability. If any part of these Terms is found unenforceable, the rest remains in effect. No waiver. Not enforcing a provision is not a waiver of it. Assignment. You may not assign these Terms; the MIT License travels with the code. No relationship. Nothing here creates a partnership, agency, fiduciary, or advisory relationship between you and the Authors. Disputes. Before pursuing any formal claim, contact the project through its public repository or package page and allow 30 days for an informal resolution. Any dispute shall be resolved on an individual basis; to the extent permitted by law, you waive participation in class or representative actions against the Authors. Language. These Terms are written in English; translations are provided for convenience and the English text prevails.
One sentence version: this is free, unregulated, open-source software that trades real money with real leverage under your own key. If you are not prepared to lose the funds in the connected account, do not connect it.