Vimana

Model-agnostic coding agent harness. Run any LLM as a coding agent with tools, skills, MCP servers, and ensemble mode.

Features

• 23 models via OpenRouter — DeepSeek, Qwen, Kimi, Gemini, Llama, MiniMax, and more
• 6 core tools — Read, Write, Edit, Bash, Glob, Grep (same interface as Claude Code)
• Skill system — Agent Skills open standard compatible, progressive disclosure, slash commands
• MCP support — Connect stdio MCP servers via .mcp.json
• Ensemble mode — Multiple expert models tackle the same task in parallel, consolidator synthesizes the best fix
• Subagent delegation — Spawn background/foreground sub-agents for parallel work
• Benchmarks — Synthetic tasks, GitHub repos, skill evaluation, subagent coordination

Quick Start

# Install
npm install
# or
bun install

# Interactive REPL
bun run src/cli.ts -m kimi-k2.5

# Non-interactive
bun run src/cli.ts -m qwen3.5-397b -p "fix the bug in server.ts"

# Ensemble mode (3 experts → 1 consolidator)
bun run src/cli.ts -e "kimi-k2.5,deepseek-v3.2,qwen3-coder" -m kimi-k2.5 "fix the race condition"

API Key

Vimana reads the OpenRouter API key from macOS Keychain:

security add-generic-password -s "openrouter-api-key" -a "$USER" -w "sk-or-..."

CLI

vimana [options] [prompt]

Options:
  -m, --model <name>       Model alias (default: kimi-k2.5)
  -p, --print              Non-interactive: run prompt and exit
  -e, --ensemble <models>  Ensemble mode: comma-separated expert models
  -c, --continue           Continue most recent session
  -r, --resume <id>        Resume session by ID
  -V, --verbose            Show full tool results
  --max-turns <n>          Max agent turns (default: 50)
  --list-models            List available models
  --list-sessions          List saved sessions

REPL Commands:
  /cost       Token usage and cost
  /model      List models
  /skills     List available skills
  /clear      Clear screen
  /help       Help
  /quit       Exit

Architecture

src/
├── core/
│   ├── agent.ts        — Agent loop (prompt → LLM → tools → execute → feed back)
│   ├── client.ts       — OpenRouter client via openai SDK
│   ├── models.ts       — 23 models with verified OpenRouter IDs
│   ├── ensemble.ts     — N experts + consolidator architecture
│   ├── skills.ts       — Skill loader, tool, slash commands
│   ├── mcp.ts          — MCP client (stdio servers from .mcp.json)
│   ├── prompt.ts       — System prompt builder (project context + repo map)
│   ├── context.ts      — Context compaction via LLM summarization
│   ├── permissions.ts  — Tool permission system (auto/ask/deny)
│   ├── hooks.ts        — PreToolUse/PostToolUse hooks
│   ├── session.ts      — Session persistence (~/.vimana/sessions/)
│   └── types.ts        — Core types (ToolDefinition, ToolContext, etc.)
├── tools/              — Read, Write, Edit, Bash, Glob, Grep + Subagent + Tasks
├── skills/             — 10 built-in skills
└── bench/              — Benchmark suites (synthetic, github, skill, subagent)

Benchmark Results

Synthetic Tasks (5 tasks — real bugs from production systems)

| Model | Score | Tokens | Cost | Avg Time | |-------|-------|--------|------|----------| | Qwen3.5-397B | 5/5 (100%) | 102K | $0.053 | 34s | | Kimi K2.5 | 5/5 (100%) | 88K | $0.051 | 49s | | DeepSeek V3.2 | 5/5 (100%) | 363K | $0.093 | 191s |

GitHub Tasks (13 real-world repo fixes)

| Model | Score | Cost | |-------|-------|------| | Kimi K2.5 | 11/13 (85%) | $0.64 | | Qwen3.5-397B | 9/13 (69%) | $1.31 | | MiniMax M2.5 | 9/13 (69%) | $0.60 |

Ensemble Mode

Multiple experts run in parallel on isolated workspace copies. A consolidator reviews all attempts and implements the best fix with full tool access.

Expert 1 (kimi-k2.5)    ─┐
Expert 2 (deepseek-v3.2) ─┼──→ Consolidator (kimi-k2.5) ──→ Final fix
Expert 3 (qwen3-coder)  ─┘

License

MIT