# Kornial

> Find and validate trading strategies. An AI agent writes a trading strategy, backtests it in an isolated sandbox on real tick data at fixed costs, and validates it on data it never touched; the person states the question and judges the result.

Kornial lets the assistant research a trading idea end to end: it writes the strategy as versioned Python files for NautilusTrader, checks and backtests it in an isolated sandbox on the instruments and date ranges the server lists, reads the stored reports, compares backtests, and applies fixed validation rules (a baseline, parameter sensitivity, a holdout on data no backtest touched) before it calls anything an edge. A viewer replays a stored backtest on a chart. It is for people researching systematic strategies. It places no trades, connects to no broker and gives no investment advice; a backtest is not evidence of future profitability.

## Connect

MCP server (streamable HTTP, OAuth sign-in in the client): https://mcp.kornial.ai/mcp

### ChatGPT

1. Turn on developer mode: Settings › Apps › Advanced settings (a paid plan).
2. Settings › Apps › Create: the server URL, authentication OAuth, no client ID or secret.
3. Connect, sign in, then pick Kornial with @ in a new chat.

### Claude

1. Customize › Connectors › Add custom connector.
2. Paste the server URL; leave the OAuth client ID empty.
3. Connect and sign in, then ask for a backtest in any chat.

### Claude Code

1. Add the marketplace: `/plugin marketplace add Secret-Algos/kornial`
2. Install the plugin: `/plugin install kornial@kornial`
3. Sign in: `/mcp`

### Codex

1. Add the marketplace: `codex plugin marketplace add Secret-Algos/kornial`
2. Install the plugin (the tools and the backtest-research skill): `codex plugin add kornial@kornial`
3. Start a new session and sign in when Codex asks for the server.
4. Tools only, without the skill: `codex mcp add kornial --url https://mcp.kornial.ai/mcp && codex mcp login kornial`

### Onboarding prompt

Paste this into your agent:

```text
Connect to Kornial, the MCP server where you write, check, backtest and validate trading strategies for me.

Server: https://mcp.kornial.ai/mcp (streamable HTTP; sign in with OAuth in the client, no API key).
- ChatGPT: Settings › Apps › Create, paste the server URL, authentication OAuth.
- Claude: Customize › Connectors › Add custom connector, paste the server URL.
- Claude Code: /plugin marketplace add Secret-Algos/kornial, then /plugin install kornial@kornial, then /mcp to sign in.
- Codex: codex plugin marketplace add Secret-Algos/kornial, then codex plugin add kornial@kornial, then sign in when a new session asks (tools only: codex mcp add kornial --url https://mcp.kornial.ai/mcp && codex mcp login kornial).

If your client needs a restart to load the tools, tell me and wait. Then call get_research_capabilities, read the strategy guide it returns, and ask me what to test. The loop is create_strategy, check_strategy, run_backtest, get_backtest, get_backtest_report, assess_strategy; quote the score assess_strategy computes, never your own judgement. Docs: https://kornial.ai/llms.txt
```

## Tools

- `get_research_capabilities`: Call first: readiness, each instrument's data coverage, costs, limits and the strategy guide.
- `list_strategy_examples`: An index of runnable example strategies to start from; one by id with its files and hypothesis.
- `list_strategies`: The user's strategies with their scores, newest first.
- `read_strategy`: Before a change: a strategy's description, notes, versions, files and latest backtests.
- `create_strategy`: A new strategy and its version one, the hypothesis in its description; nothing runs.
- `write_strategy`: Replace a strategy's notes, or its description until its first backtest freezes it, or archive it (nothing is deleted).
- `check_strategy`: Before every backtest of changed code: import and configure it in the sandbox; smoke also replays one day.
- `run_backtest`: Start a backtest of a version on one or more instruments and a date range, or of several versions on one account; returns its id at once.
- `get_backtest`: Follow a backtest: its status, then its headline figures or the failure's file and line.
- `cancel_backtest`: Stop a running backtest, or every running cell of a study, when the user asks.
- `list_backtests`: Backtests by strategy, version, status or time.
- `get_backtest_report`: One stored report of a succeeded backtest (robustness, breakdown, log, chart, tables), a page at a time.
- `compare_backtests`: 2 to 5 backtests side by side; comparable is false for unlike conditions.
- `assess_strategy`: A version's score from its stored backtests, 0 to 100, with every figure it read; quote it, never your own judgement.
- `run_study`: After the in-sample backtest: a walk-forward over 2 to 6 folds, or a small grid or random search; each cell is an ordinary backtest and counts as a trial.
- `get_study`: Follow a study as it runs its cells: the fold table with its efficiency, or the trials table with the best cell's deflated Sharpe.
- `prepare_upload`: From a local shell: an upload URL for a strategy folder (the plugin's send script uses it).
- `analyze_market`: Before writing: an instrument's volatility, trend or reversion, active hours and correlations, from the stored candles.

## Docs

Each chapter as markdown, public; all of them in one file at https://kornial.ai/llms-full.txt. An agent connected over MCP reads the same guide from `get_research_capabilities`, with the exact instrument ids.

- [Get started](https://kornial.ai/docs.md): Your agent writes a strategy, backtests it in a sandbox at fixed costs and validates it on data it never touched; you state the question and judge the result. These pages are for you; your agent reads the same guide through the tools.
- [Connect your agent](https://kornial.ai/docs/connect.md): One MCP server for every client. Sign in once from the client; the tools act on your strategies and backtests.
- [Strategy guide](https://kornial.ai/docs/guide.md): The text your agent reads: how the loop works, what the engine accepts, how a result is validated, every indicator by name.
- [Methodology](https://kornial.ai/docs/methodology.md): What a backtest counts as, the four checks, the edge rules, the verdicts and the score: computed by the server from stored backtests, never by the agent.
- [Data & engine](https://kornial.ai/docs/data.md): What a backtest can read, what it pays, how its orders fill and how far it can go: get_research_capabilities, as of now.
- [Examples](https://kornial.ai/docs/examples.md): Worked strategies to start from, each with the hypothesis it tests.

## More

- [Website](https://kornial.ai): sign in, your strategies and backtests, the replay of each backtest.
- [Support](https://kornial.ai/support): info@kornial.ai
- [Privacy](https://kornial.ai/privacy), [Terms](https://kornial.ai/terms)