For LLMs

Two ground-truth artifacts you can paste straight into Claude, GPT, Gemini, or any other assistant so it answers Flint questions from the canonical docs.

Drop-in URLs

File	Use it for	Link
llms.txt	A short, structured manifest with absolute links to every doc page + the proto. Follows the llmstxt.org convention — paste the URL into a tool that fetches it (Claude Projects, ChatGPT custom GPTs, Cursor docs).	/llms.txt
llms-full.txt	Every page on this site concatenated into a single Markdown file (~1500 lines). Drop the contents directly into your model's context to ground completions in the docs without any fetching.	/llms-full.txt
api.proto	The bundled protobuf schema for every service in flint.spot.v1.	/api.proto
openapi.yaml	OpenAPI 3.1 schema for the gRPC-Web service URLs.	/openapi.yaml

Build a maker bot

Use this generator to create a prompt for an assistant that can build a rudimentary maker bot with the Flint SDK. Pick the implementation language and quoting strategy, then copy the prompt into the coding agent.

TypeScript maker bots

The TypeScript SDK is useful for market data, stats, auth, and MakerService reads, but it does not expose transaction signing/submission for quote updates. Use Rust or Python for a bot that directly quotes a market.

Maker Bot Prompt

Generate a task prompt for an assistant that can read the Flint docs and build a first maker bot.

LanguageRustPythonQuoting strategyOracleOffsetOrderListLinearDistributionMarket

You are building a minimal but real Flint maker bot. Write your own small bot; do not vendor an example repo unless explicitly asked. Source order: - https://docs.flintlabs.dev/llms.txt - https://docs.flintlabs.dev/llms-full.txt Use the public docs above as the source of truth. If external fetching is unavailable, use llms-full.txt content supplied in the agent context. Target: - Language: Rust - SDK/package: flint-api-client - Install/dependency guidance: - Add the crate from the Kellnr registry in Cargo.toml: flint-api-client = { version = "1.6", registry = "kellnr" }. - Point Cargo at the registry in ~/.cargo/config.toml with: [registries.kellnr] index = "sparse+https://kellnr.xantasoft.com/api/v1/crates/". Pulling needs a read-only registry token — store it with: cargo login --registry kellnr. - SDK entry points to use: Client::builder(), authenticate(), start_quoting_core(), QuoteBuilder - Market to quote: SOL/USDC - Market identifier note: when using Rust QuoteBuilder, pass the base token identifier (SOL) rather than the full pair string (SOL/USDC); use the pair only for user config and market-data RPCs. - Quoting strategy: OracleOffset - Strategy summary: Quote a single two-sided bid/ask around an external fair price. This is the recommended default for a first real maker bot. Concrete environment defaults: - Default network for a runnable demo: devnet. Make mainnet a config switch. - Devnet public API: https://devnet.api.flint.trade - Devnet authed API: https://devnet.makerapi.flint.trade - Devnet Solana RPC: https://api.devnet.solana.com - Devnet program id seed from the checked-in SDK lookup-table config: ssuJqL67UPz8vSZf2Rntp8iuSFCe1hy9GFNrdZEoJWs. Still require FLINT_PROGRAM_ID config and verify it against the market catalog and Solana RPC before submit mode. - Mainnet public API: https://mainnet.api.flint.trade - Mainnet authed API: https://mainnet.makerapi.flint.trade - Mainnet Solana RPC default: https://api.mainnet-beta.solana.com - Default market: SOL/USDC - Default fair source for OracleOffset: external fair first. For SOL/USDC, prefer Pyth Hermes when available; allow a static fair for tests. Do not implement book-mid as a source — with no usable external fair, pull quotes. Build a small CLI or long-running process with configuration for keypair path, network/endpoints, trusted Flint program id, Solana RPC URL, market selection, fair source, quote size, spread/band settings, target inventory, soft inventory limits, tick interval, dry-run vs submit mode, and kill-switch settings. Assume the operator has already completed Flint onboarding for the target network — maker_id created, the bot's hot quoter pubkey registered/delegated as a quoting authority, and deposits, quote-side caps, and global/cross-spread params configured for the market. Do not ask the operator to confirm any of this and do not bake confirmation prompts into the bot; assume it is set up correctly and let real failures surface on their own (authentication fails clearly if the key is not a registered quoter; program-id verification fails on mismatch). The only operator input the bot needs is ordinary configuration such as which network to target. Hard invariants: - Follow the public SDK source and examples exactly where API names differ from memory. - Do not invent protocol fields, RPCs, or helper names. If an SDK type or method is uncertain, inspect the public repository or public docs first. - Load only a registered on-chain quoter keypair for submit mode. This hot key signs authentication and quote transactions; the admin wallet should not be required at runtime. - Authenticate once at startup and fail clearly if the key is not registered as a quoting authority. - Assume the maker_id and the loaded quoter pubkey are already set up correctly; rely on the startup authentication check to fail clearly if the key is not a registered quoting authority, rather than asking the operator to confirm. - Discover or validate the target market before quoting. Prefer catalog metadata over hardcoded decimals, lot sizes, or spot ids. - Verify the program id against trusted config before submitting quotes. - Include a first-run market setup path for what the quoter key can set: enable=true and a global cross-spread entry for quote spot 0 with zero added spread. Do not claim the bot can configure deposits, soft_max_balance caps, or quote-side USDC caps; those are admin/dashboard prerequisites (set via the deposit/withdraw path, not update_quoting_params) unless you find and name a real SDK/API path. - If you choose read-then-verify setup instead of always setting quoter-controlled params, name the exact source of truth used for each check. Otherwise assume the admin prerequisites are already in place rather than gating submit mode on operator confirmation. - State the fill-gating warning once in README/config docs: accepted quotes may not fill unless the maker is enabled, inventory is deposited, quote-side caps are set, and global/cross-spread params are configured. - Publish one small two-sided quote around fair price, then loop and refresh on fair changes, timer ticks, and inventory changes. - Inventory changes must affect quoting: while below the soft limit, skew the fair/anchors away from excess inventory; once the soft inventory limit is reached, pull all quotes. The quote is either two-sided or fully pulled — do not implement one-sided (bid-only/ask-only) shapes. - Pull quotes on graceful shutdown, kill-switch, stale/outage fair feed, chain-tip outage, program-id mismatch, and configured risk-limit breaches. - Map every commit and pull to one outcome — confirmed, unknown (timeout/disconnect/lag), or failed — and log it clearly. Treat unknown as "on-chain state may differ" and reconcile before re-sending; do not collapse unknown into failed. - Never run two quoting processes for the same maker_id. Keep the implementation simple. Deliberately do NOT add: - A multi-level ladder or depth. Quote one level per side; no extra-levels / level-step config. - One-sided shapes. The quote is either two-sided or fully pulled — skew hard while below the soft inventory limit, pull everything at the limit. - A balance/fills subscription, spawned task, or watch/broadcast channel for inventory. Poll maker balance once per tick instead. - Book-mid as a fair source. Use external (Pyth) and static only; with no usable external fair, pull quotes. - A slot-advance watchdog. Treat chain-tip health as "is there a usable blockhash/chain tip?" and pull if not. - Multiple outcome types. Use one outcome — confirmed, unknown, or failed — for both commits and pulls. - The dry-run vs submit fork sprinkled through the code. Compute the quote as a pure plan, then have one apply step that either logs it (dry-run) or submits it; keep the fork in that one place. Do keep (load-bearing, do not simplify away): - Install params once, then send only fair updates on later ticks when the quote shape/risk params are unchanged. Re-sending full params every tick resets accumulated on-chain risk state. - Treat timeout/disconnect/lag and post-signing errors as unknown (not failed), and reconcile intended vs on-chain state before replacing quotes. Language-specific references: - Use the default quoting feature; it includes QuoteBuilder, QuotingCore, transaction packing, signing, and receipts. Language-specific requirements: - Load the registered Solana quoter keypair, build a Client with wallet(Arc<Keypair>), authenticate once, then start_quoting_core with the same keypair. - If the CLI accepts BASE/QUOTE such as SOL/USDC, parse it for user ergonomics but pass only the base token identifier such as SOL, a base spot id, or a base mint to QuoteBuilder. The Rust quoting builder resolves one base spot and uses Flint's global quote spot internally. - Use human prices and sizes with Rust quoting builder types. Do not add manual unit lowering around OffsetSpec, LinearParams, ParamsUpdate, or order-list helpers; the Rust builder lowers them internally. - Do not persist application-managed oracle/order sequence counters for Rust. QuotingCore seeds per-spot oracle_sequence, order_sequence, and client_order_id from nanoseconds since epoch. The rule you still must enforce is one live quoting process per maker_id. - For setup, the quoter can set enable=true and cross-spread entries. soft_max_balance caps (both sides), quote-side USDC caps, and deposits are admin/dashboard prerequisites set via the deposit/withdraw path, not update_quoting_params; document them and do not claim the bot can set them with the quoter key. - Commit each tick with QuoteBuilder and await at least accepted(); use landed_within for the initial install, shutdown pulls, and health checks. - Poll maker balance once per tick (get_balance) and recompute the quote from it. Do not subscribe to a balance or fills stream for a first bot; a per-tick poll is adequate at a multi-second tick. - Keep runtime state in one struct with &mut self methods (e.g. context, quoting core, and loop state together) rather than threading several context/state objects through every function. - On shutdown or kill-switch, explicitly withdraw live quotes. Prefer disabling the market with ParamsUpdate enable(false) for continuous strategies or cancel_all for OrderList, then await the receipt before exiting when possible. Strategy-specific requirements for OracleOffset: - Default the fair source to an external reference, preferably Pyth Hermes for supported assets; allow a static fair for tests. Do not implement book-mid as a source — with no usable external fair, pull quotes. - Install the quote and risk params once, then on later ticks send only updated fair prices when the quote shape/risk params have not changed. Re-sending full params every tick resets accumulated on-chain risk state, so this distinction is load-bearing — keep it. - Use post_only=false initially and a staleness window around 5 slots. - Quote one level per side, not a multi-level ladder. - Apply inventory skew: shift bid and ask anchors away from the side that is accumulating inventory relative to a target balance while below the soft limit, and pull all quotes once the soft inventory limit is reached. Do not implement one-sided (bid-only/ask-only) shapes — the quote is either two-sided or fully pulled. - Pull quotes immediately on fair-feed staleness/outage, program-id mismatch, chain-tip outage, hitting soft_max_balance, or an explicit kill-switch. - Add simple reconciliation: after receipts, track intended quote state and log/refuse to continue when a stale sequence or unknown outcome means on-chain state may differ. This timeout=unknown then re-send/reconcile rule is load-bearing — keep it. - Use a single OffsetSpec level per side in human units. price_offset is the quote-asset distance from fair; size is a human base-token amount. Use a staleness window such as 5 slots. Do not build a multi-level ladder. - Use QuoteBuilder::new().oracle_offset(base_symbol, |b| b.with_spread(...).with_fair((mid, mid))).commit(&mut core). On later ticks, use with_fair((new_mid, new_mid)) only when spread/risk did not change. - For the global USDC book, configure cross-spread with ParamsUpdate::new().enable(true).with_cross(0_u16, Some(0.0)). Some(0.0) means a zero added-spread global entry; None disables crossing into that peer. Acceptance criteria: - Verify with cargo fmt and cargo test or cargo check. If no live endpoint/keypair is available, stop at compile plus unit tests for pricing/risk logic and make submit mode require explicit config. - Include unit tests or isolated pure-function tests for fair selection, spread/band calculation, inventory skew, risk kill-switch decisions, and shutdown quote-pull construction where reasonable. - If live devnet credentials/endpoints are available, add a dry-run or submit-mode smoke path and document the exact command. If they are not available, define done as compile/byte-compile plus pure logic tests. Deliverables: - A working project layout with dependency files and source code. - A README that explains configuration, setup prerequisites from Flint onboarding, dry-run behavior if implemented, and the command to run the bot. - Keep the first version intentionally simple: one market, one maker authority, one strategy, one quote level per side, external fair (pull when unavailable), inventory skew, quote-pull shutdown, and defensive error handling.