Mirkaso
Built at Meta Global AI Developer Hackathon · Oct 26, 2026 · Remote

Mirkaso Copilot is an agentic market assistant built on Meta's Muse Spark (muse-spark-1.3) via the Meta Model API. It runs inside Mirkaso (mirkaso.com), a live, bootstrapped crypto derivatives analytics platform with paying subscribers, built and run by a solo founder. THE PROBLEM Derivatives data (open interest, funding, scanner signals, sector rotation, macro regime) is spread across many screens. Answering a simple question like "which sector is overheating and why?" means opening several tabs and combining the numbers by hand. Most retail traders can't do that. WHAT IT DOES A trader asks a question in plain language. Muse Spark decides which tools to call, calls several in parallel, and writes a short answer based only on the results. The 4 tools read live Mirkaso data: - Scanner signals (anomaly scoring over the top-150 USDT perpetual pairs) - Sector indices - Macro regime score and history - Coin derivatives (price, open interest and their changes) HOW WE KEEP IT HONEST - Structured answers: numbers are resolved server-side from tool output, so the model doesn't type them freely. - A numeric grounding guard checks the answer against tool results. - Every answer shows its sources and an as-of time. - If the data isn't available (for example a long/short ratio), Copilot says so instead of guessing. - The existing output guard against advice-style wording stays in place. - Tier-based gating and rate limits are decided by the host app, never by the model. BUILT FOR PRODUCTION Copilot is live on mirkaso.com/app behind subscription-tier limits; one answer costs about 2 cents. The agent loop has step, token and timeout limits, duplicate-call detection and structured errors that never leak internals. An EU forwarding proxy handles Meta API access from Europe. ARCHITECTURE Hybrid: the Copilot package (tools, Meta client, agent loop, grounding guard, demo CLI) is open source in this repo (MIT). It reads data through a DataProvider interface. The private Mirkaso product implements that interface, so the open package carries no business logic or secrets. TIMELINE (declared honestly) Before the hackathon: the Mirkaso platform itself (closed source), and an opt-in Meta provider in the private repo committed on Oct 26 at 12:32 UTC, before the kickoff ended. During the hackathon: this repository was created on Oct 26 at 18:38 UTC, after the kickoff, and all of its commits come later. In the private product repo from Oct 27: the data adapter, Copilot API behind a feature flag, chat panel, and production release. NEXT More tools (long/short ratios, ETF flows, options), multilingual answers, and an evaluation set that compares Muse Spark with our production model on accuracy, latency and cost.