MINARAWhat is Minara Agent
What Minara Agent is, what problems it solves, and how it differs from other AI agents
Minara Agent
👥 This section is for: everyone deciding whether Minara fits their use case. Plain-language summary first, then the technical differentiators. Unfamiliar acronym? The Glossary has 1-line definitions for everything on this page.
In one sentence: Minara is a chat-first AI assistant that does the same things you'd do across a dozen trading and research apps — price checks, portfolio balances, swaps, perps, research reports, alerts — except you describe what you want in plain English and the agent figures out which data source, exchange, or tool to use.
Minara Agent is an open-source TypeScript AI agent purpose-built for digital finance. It trades spot crypto, perpetual futures, DeFi, equities, commodities, and FX from a single runtime with one memory, one safety layer, and one learning loop.
On-chain data, DeFi protocols, and cross-chain markets are first-class data sources alongside equities and commodities. Minara is built for the convergence of AI, blockchain, and personalized finance.
Built with a custom LLM agent loop, an in-house skill and tool registry (no MCP for internal tools), SQLite-backed state (WAL + FTS5), a sandbox-first execution model, persistent memory with financial personalization profiles, and a self-learning engine that promotes signal methodologies from real trade outcomes.
Key numbers
| Metric | Value |
|---|---|
| Builtin skills | 40+ |
| External skills | 33 (CoinGecko, DefiLlama, OKX, Polymarket, ...) |
| Tools | 200+ |
| Scenarios | 33 (25 primary + 8 composable) |
| Permission tiers | 4 (READ_ONLY → MANUAL_ONLY) |
| State backend | SQLite (single file, WAL + FTS5) |
| Stack | Node ≥ 24, TypeScript 5.7, ESM |
What problems it solves
-
Multi-market coverage from one agent. Most trading bots are single-market. Minara handles spot, perps, DeFi, stocks, and prediction markets through one conversation, one memory, and one audit trail.
-
Safety for fund-moving AI. Every tool carries a permission tier. Fund-moving calls simulate first, show the resolved order, and wait for explicit confirmation. An emergency stop halts everything instantly. Six-stage trade safety: token validation, position sizing, exposure limits, slippage protection, daily caps, and stop-loss monitoring.
-
An agent that improves over time. The learning loop records successful tool sequences, the role-memory system reflects on decision outcomes, and methodologies graduate via Wilson confidence. The agent gets faster and more accurate at recurring tasks without model fine-tuning.
-
Intent-aware routing. The L0.5 scenario classifier maps user queries to procedural playbooks in sub-100ms, preloading the right skills before the LLM runs. 33 scenarios cover market analysis, trading, research, verification, and more.
How Minara Agent Differs from General-Purpose Agents
 OpenClaw |  Hermes Agent |  Minara Agent | |
|---|---|---|---|
| Purpose | Self-hosted gateway: connect any chat platform to AI coding agents | Autonomous self-improving agent for coding and research tasks | Purpose-built for digital finance. Every design choice traces back to safety, accuracy, and compounding knowledge. Finance demands correctness — a wrong answer is not just unhelpful, it costs real money |
| Built-in tools & skills | Coding agents via connected AI backends (Claude Code, Gemini CLI, etc.); no built-in domain tools | General coding and research tools; MCP server support for extensibility | Standard toolkit (web search, code execution, file ops, terminal, MCP) plus 40+ finance-specific skills covering market analysis, on-chain data, perps, DeFi, and deep research |
| Intent routing | Channel allowlists and per-agent workspace isolation; routing by mention pattern | Spawns isolated subagents for parallel workstreams; collapses multi-step pipelines via programmatic tool calling | Pre-AI classifier routes each turn into 1 of 41 finance-specific playbooks before any tool runs |
| Memory | Per-sender and per-workspace session context; no cross-session learning | Agent-curated memory with periodic reflection; dialectic user modeling that deepens across sessions | Domain-scoped memory captures analysis patterns, trade rationale, and real outcomes across sessions |
| Self-learning | Static capability set; no autonomous improvement | Autonomous skill creation and self-improvement during use; FTS5 cross-session recall with LLM summarization | Analytical rules emerge from real trading performance, graduate through a statistical confidence gate, and auto-demote when accuracy drops |
| Execution security | Channel allowlists, sender-level access restrictions | Command approval workflows; container isolation for untrusted code execution | Sandboxed filesystem, kernel-level subprocess guard, and AI-invisible credentials form 3 interlocking layers |
| Deployment | Self-hosted on your own machine; optional remote via Tailscale or SSH | Cloud-native on any infrastructure from a $5 VPS to serverless; hibernation support | 1 Docker container with SQLite only. No Postgres, Redis, or vector database. Local and production are identical |
| Fund-moving safety | Not applicable | Not designed for fund-moving | 2-step confirmation for every fund move. 4-tier permission model with a per-conversation risk ceiling. Automation bypass is CI/testing-only |
| Financial profile | Not applicable | Not applicable | 10 behavioral dimensions (risk appetite, markets traded, finance knowledge, decision style) continuously re-inferred. Explicit settings always take priority |
| Trading memory | Not applicable | Not applicable | Captures real trade rationale and outcomes across sessions. Upload your trade history to build genuine habits, not assumed ones |
Design philosophy
Minara is built around three convictions:
-
An agent that is occasionally unsafe is unusably unsafe. Safety is structural: permission tiers are enforced by hooks, the sandbox resolver blocks traversal at the syscall layer, and the confirm gate is the only process-wide bypass. See Design Philosophy.
-
An agent that learns from real trading outcomes. Methodologies and scenario playbooks emerge from actual trade results, graduate through a statistical confidence gate (Wilson lower bound ≥ 0.55), and auto-demote when accuracy drops. Domain-scoped memory preserves analysis patterns and trade rationale across sessions. The agent improves with every real result — no model fine-tuning required. See Self-Improving Agent.
-
Skills are lazy-loaded prompt fragments, not agents. A skill is a declarative bundle (prompt + tool whitelist + routing metadata) that the L0 router scores deterministically before the LLM sees the catalog. Two-thirds of the activation pipeline is testable without an LLM. See Skills System.