Opsora routes any task to the best LLM, MCP tool, and skill across 8+ providers and 40+ models — then executes it autonomously with a dual-model orchestrator. One terminal, one API, zero lock-in.
$ opsora "refactor the auth middleware, then deploy" ✓ intent:code → tokenhub/kimi-k2.7-code ✓ plan: 4 subtasks · planner=reasoning · explorer=nemotron-49b → mcp__github__create_branch …ok → skill__test-driven-development …ok → skill__systematic-debugging …ok ✓ 2 files edited · 6 tests green · $0.0021 spent $
Opsora is a modular monolith that unifies routing, tools, memory, and orchestration behind one consistent interface.
Intent-aware, capability-aware routing picks the right model for code, vision, reasoning, or speed — with automatic fallback across providers.
A reasoning planner pairs with a fast explorer model to plan, execute, and verify multi-step work autonomously.
GitHub, Supabase, Gmail, filesystem, SQLite, NVIDIA NGC, Discord, Telegram, Google — plus browser & data-source bridges.
A plugin system and a skill bridge expose superpowers workflows and kimi plugins as first-class agent capabilities.
SQLite-backed memory, sessions, and a code knowledge graph with FTS5 keep context across every run.
Real-time token and spend tracking per model, with cost- and speed-preference routing baked in.
Bring your own keys. Opsora speaks OpenAI-compatible, AWS Converse, and Kimi protocols, so you can mix free-tier and flagship models freely.
A tight loop that plans, acts, verifies, and reports — no hand-holding required.
Regex + capability scoring picks the right model and tools for the intent.
The reasoning planner decomposes the request into bounded subtasks.
A ReAct loop runs tools, MCP calls, and skills across parallel sub-agents.
A separate verifier model confirms completion and tracks cost before reporting.
Install, drop in your keys, and route your first task.
pip install opsora-cli opsora