Local-first  •  LM Studio · Ollama · llama.cpp · MCP

Claude Code — but local.

Frontier-grade coding agents that run entirely on your machine. Hardware-aware parallel sub-agents, intelligent model routing, zero cloud cost. Your code never leaves the laptop.

Install — npm i -g lmstudio-ollama-mcp Read the docs →
$ lmstudio-ollama-mcp doctor — diagnose providers & hardware
$ lmstudio-ollama-mcp models — list GGUF / Ollama models
$ lmstudio-ollama-mcp "add tests for src/utils/logger.ts" (alias: forge)
↳ gemma-3-12b-qat · 6.5GB Q4_0 · 8 agents · 1.2 s — local only, no API key
You are an autonomous coding agent. Set up lmstudio-ollama-mcp in this repository and start working: 1. Install: npm install -g lmstudio-ollama-mcp ...
Show full prompt +
You are an autonomous coding agent. Set up lmstudio-ollama-mcp in this repository and start working: 1. Install: npm install -g lmstudio-ollama-mcp 2. Verify: lmstudio-ollama-mcp doctor && lmstudio-ollama-mcp models - Expect: LM Studio at http://localhost:1234/v1 or Ollama at http://localhost:11434 - If no models, guide the user: LM Studio -> Load model -> Start Server, or: ollama pull qwen2.5-coder:7b 3. Init: lmstudio-ollama-mcp init (creates lmstudio-ollama-mcp.json; supports hybrid via OPENAI_API_KEY / ANTHROPIC_API_KEY) 4. Work: lmstudio-ollama-mcp "explore the codebase, then implement <TASK> and verify with tests" - Hardware-aware parallel sub-agents run automatically (--parallel auto based on CPU/RAM) - Use --model lmstudio:<model> or --provider ollama to force local Keep edits minimal, verify with bash/tests, summarize changes.
LM Studio Ollama llama.cpp GGUF Apple Silicon CUDA · Vulkan
Features

Everything you expect from Claude Code / Codex — re-engineered for on-device models. No vendor lock-in, no per-token bill.

Local-first engine

Claude Code loop,
on your hardware

Read → plan → edit → verify with 7 sandboxed tools: read, write, edit, bash, glob, grep, list. Exact-string edits, capped iterations, truncated safety.

↗ Learn more
Parallel sub-agents

Hardware-aware
fan-out

Tasks decompose into 2–6 sub-agents. Scheduler caps concurrency by cores × RAM. Apple Silicon unified-memory bonus. Up to 16 parallel on 64 GB.

↗ Learn more
Intelligent router

Frontier plans,
local executes

Auto strategy: trivial/small → local 7B, medium/large → frontier. Prefer-local list: lint, format, test, search, explain. Save 90% of API cost.

↗ Learn more
All local runtimes

LM Studio · Ollama
llama.cpp · any GGUF

OpenAI-compatible API. LM Studio filesystem scan, Ollama native /api/tags, llama.cpp health check. Reasoning-content normalized for Gemma/Qwen.

↗ Learn more

Ship faster — without the cloud.

One binary, zero daemons. lmstudio-ollama-mcp (alias forge) respects your filesystem, Git history, and privacy. Bring your favorite GGUF.

Get your free setup
lmstudio-ollama-mcp — local-first coding agent
$ lmstudio-ollama-mcp --parallel 4 "migrate codebase from Jest to Vitest"
▶ Decomposed into 3 sub-tasks:
  • t1 Explore & plan   [search → lmstudio/gemma-3-12b]
  • t2 Implement    [code → lmstudio/qwen3-27b]
  • t3 Verify      [test → lmstudio/gemma-3-12b]

✔ t1  search   3.1s  glob + grep + read — 4 files relevant
✔ t2  code     8.4s  edit_file x3, write_file x1
✔ t3  verify   5.2s  bash: vitest run — 24 passed

Synthesis: Migrated Jest → Vitest (config, mocks, 1 test helper).
Files: vitest.config.ts, package.json, tests/*.test.ts
Workflows

Real tasks, local execution. No API keys unless you want hybrid routing.

For a 12k LOC TypeScript API, lmstudio-ollama-mcp refactored error handling across 18 files in parallel — 8 sub-agents on an M3/16GB, all local. No data left the machine.

Learn more ↗

For a B2B SaaS frontend, it generated missing Vitest coverage (+34%) by routing small tasks to a local 7B and synthesis to frontier — 90% cheaper.

Learn more ↗

For a CLI tool repo, it migrated logging, added a hardware detector, and opened a PR — entirely via lmstudio-ollama-mcp "…" from a laptop on a train.

Learn more ↗

Install in 30 seconds

bash
# 1 — install
npm install -g lmstudio-ollama-mcp
# aliases: forge, forgecode — same binary

# 2 — start your local runtime (one of)
lms server start          # LM Studio
ollama serve              # or Ollama

# 3 — verify
lmstudio-ollama-mcp doctor
lmstudio-ollama-mcp models # or: forge doctor / forge models

# 4 — go
lmstudio-ollama-mcp "refactor src/providers into a registry"
lmstudio-ollama-mcp --parallel 4 "add tests + docs for the new registry"
Full documentation → See docs.html for detailed provider, routing, and MCP guides.
  1. 1LM Studio: Developer → Local Server → Start. Load a model (Gemma 12B QAT or Qwen3 27B recommended on 16GB). Auto-detected via /v1/models + filesystem scan.
  2. 2Ollama: ollama pull qwen2.5-coder:7b. Uses /api/tags for metadata.
  3. 3llama.cpp: ./llama-server -m model.gguf --port 8080. Any GGUF works.
  4. 4Hybrid (opsiyonel): OPENAI_API_KEY / ANTHROPIC_API_KEY set edin. Router zor işleri frontier'a, küçük işleri locale yönlendirir — %90 tasarruf.
  5. ↗Alias'lar: forge ve forgecode aynı binary — kas hafızası korunur. forge doctor = lmstudio-ollama-mcp doctor.