Spark — local-first agentic coding in your terminal

Spark is an agentic coding assistant that runs entirely against your own local LLM server. No cloud, no third-party API keys, no code leaving your machine — point it at an OpenAI-compatible endpoint (built for vLLM serving Qwen3.6, and it adapts to other local models) and it reads, edits, and runs your project with your approval.

By Solderspark.

Install

npm install -g @solderspark/spark-code

That gives you the spark command. (Node 20+.)

Quick start

1. Run a local model server that exposes an OpenAI-compatible API. The easiest is Ollama:

   ollama pull qwen2.5-coder

   (Or LM Studio, vLLM, llama.cpp — anything OpenAI-compatible.)
2. Run Spark — it auto-detects a running local server (Ollama / LM Studio / vLLM / llama.cpp) and adapts to whatever model it serves:

   spark

   First run walks you through setup if nothing's detected. To point it somewhere specific:

   spark --url http://localhost:8000/v1 --model qwen3.6-35b

That's it — ask a question, and Spark works in your current directory.

Modes

Switch any time with the /mode picker (or /ask, /bypass, /agentic):

| Mode | Tools | For | |---|---|---| | Ask (default) | approve each tool (y/a/n) | staying in control | | Bypass | auto-approve all tools | "stop asking, I'm watching" | | Agentic | auto-approve + run unattended | long autonomous tasks | | Plan (/plan) | plan first, then execute | risky multi-step work |

What it does

• A real agent loop — the model calls tools (Read, Grep, Glob, Edit, Write, Bash, subagents), the harness runs them, results feed back, repeat.
• Undo — /undo reverts the last turn's file changes (and deletes files it created). Every turn is checkpointed.
• Self-correction — after it says it's done, Spark runs your typecheck/build on the changed files and fixes what it broke before stopping.
• Safety nets — catastrophic shell commands (disk wipes, curl | sh, fork bombs) are hard-blocked even unattended; edits preserve a file's real encoding (UTF-8 / BOM / UTF-16); a loop guard steers out of doom-loops.
• Capability detection — Spark probes the served model and adapts (context window, tool-calling, parallel tools, reasoning), so non-qwen local models work too, with safe defaults.
• MCP — add custom/external tools (browser automation, databases, …) via MCP servers in ~/.spark-code/mcp.json or ./.spark-code/mcp.json.
• Sessions — /session save, resume with spark --resume <id>.
• Project memory & prompt layering — ./SPARK.md and file-based slash commands in ./.spark-code/commands/.

Useful flags

| Flag | Description | |---|---| | -u, --url <url> | OpenAI-compatible server base URL | | -m, --model <name> | Model id (the one your server serves) | | --think | Enable the model's thinking/reasoning for this session | | --resume <id> | Resume a saved session | | --print <prompt> | Run one turn non-interactively, print to stdout, exit (scriptable) | | --allow-all | With --print, auto-approve every tool (for CI) |

Run spark --help for the full list, and /help inside the session for slash commands. /diagnose checks your server connection and the model's capabilities.

Supported models

Tuned for qwen3.6-35b (A3B MoE, sequential tool calls) and qwen3.6-27b (dense, parallel tool calls) via vLLM, but capability detection lets other OpenAI-compatible local models work with conservative defaults. Set --model / sparkCode.model to the id your server actually serves; /diagnose confirms it.

Privacy

Everything runs against the server you configure. Spark sends your prompts and the files it reads to that endpoint only. No telemetry, no third-party service.

Also available for VS Code

Prefer a GUI? Spark is also a VS Code extension (same engine, a chat panel with inline diffs, grounded Plan mode, and a context visualizer) — search "Spark" in the VS Code Marketplace.

© Solderspark · https://solderspark.com · See LICENSE for terms.