Claude Code Ollama

You are one of the builders shortlisted from across India to compete in person in Forge Sprint 01, the inaugural fortnightly building tournament from nmg.labs. On Saturday 6 June 2026 you will spend six hours building a real, working AI-powered digital marketing automation using Claude Code at our Gurgaon office.

What is at stake:

• Champion: a structured paid NMG apprenticeship, internships and full-time opportunities with a clear performance-driven path to a software engineering career.
• Runners-up: priority fast-track invitations to apply directly to future Forge Sprints.
• All shortlisted builders: lifetime membership of the private Forge Alumni developer network and standing invitations to NMG masterclasses.

We test your ability to orchestrate AI agent loops, run parallel sub-agents, drive a local model well and ship clean client-ready reports. This is a digital marketing sprint, so you are expected to arrive with working knowledge of SEO and digital marketing fundamentals.

There is no cheat sheet on Sprint Day. Nobody will hand you the domain knowledge. You research and study it yourself before the event. Section 05 links every resource you need. Come prepared.

Forge Sprint 01 runs on one stack: Claude Code driven by an open model through Ollama. No Anthropic API key, no paid subscription, no router and no alternative coding tool. Claude Code is the agent, Ollama serves the model, and the same ollama launch claude command wires them together with web search and sub-agents built in, so you never configure an MCP server.

There are two ways to run the model, and you pick the one that fits your laptop. Both are free. Track A runs the model on Ollama's servers (best for low-RAM laptops). Track B runs the model fully on your own machine (best if you have 16 GB+ RAM). Whichever you pick, Claude Code behaves the same.

Track A · Ollama Cloud (free), for laptops under 16 GB RAM

The model runs on Ollama's GPUs, so even an 8 GB laptop works. You sign in to a free Ollama account, no card, no API key. The free tier is metered by compute time, not tokens or requests, so the trick is to use a low-usage model and spend that compute only on real file edits (see Section 06, Sprint-day strategy).

Track B · Local model (free, unlimited), for laptops with 16 GB+ RAM

The model runs entirely on your machine. Running on your own hardware is always unlimited, there is no quota at all, just RAM and speed. Pick the largest size your RAM can hold:

Why open models on purpose

These models do not do all the thinking for you, so the Sprint actually tests your skill, not the model's:

• It tests your prompting. You break a big build into small, modular prompts (schema first, then an endpoint, then a UI piece) instead of asking for everything at once. Tight prompts also save cloud compute.
• It tests your debugging. The model will occasionally produce a minor bug or a deprecated call. The builders who win read the error, understand the code and fix the model's mistakes.

Learn Claude Code and Ollama (start before Sprint Day)

• Anthropic Skilljar: Claude Code 101
• Anthropic Skilljar: Claude Code in Action
• Anthropic Academy (free courses)
• DeepLearning.AI: Claude Code, a Highly Agentic Coding Assistant
• Anthropic: Claude Code Best Practices
• Anthropic: Claude Code Docs
• YouTube: Anthropic channel
• Ollama Docs: Claude Code integration
• Ollama: web search and sub-agents in Claude Code
• Ollama Docs: Cloud models (free tier)
• Ollama: all cloud models and their usage levels
• Ollama: qwen3-coder-next model card

Do all of this at home before Sprint Day. Downloads and sign-ins on venue Wi-Fi waste your build time. Arrive with a working, verified setup. If anything fails, read Section 03b (errors) first, then email labs@nmgdigital.com.

Step 1 · Install Node.js 18+ and Git (both tracks)

If you do not already have them, install the Node.js LTS build from nodejs.org/en/download and Git from git-scm.com/downloads. Verify:

node --version    # must be v18.0.0 or higher
git --version     # any recent version is fine

Step 2 · Install a code editor (both tracks)

VS Code is recommended: code.visualstudio.com. Any editor you are comfortable with is fine.

Step 3 · Install Ollama and Claude Code (both tracks)

Install Ollama for your OS from ollama.com/download (Windows, macOS, Linux), then install Claude Code. Verify both:

ollama --version                              # confirms Ollama
npm install -g @anthropic-ai/claude-code
claude --version                              # confirms Claude Code

Step 4 · Set the context window (mandatory, both tracks)

Claude Code needs a large context window or it degrades and prints raw tool JSON instead of running tools. Set this once, then open a new terminal so it takes effect. Full details are in Section 03a.

# Windows (PowerShell), then reopen the terminal:
setx OLLAMA_CONTEXT_LENGTH 65536

# macOS / Linux:
export OLLAMA_CONTEXT_LENGTH=65536

Now do Step 5 for your track. Track A if your laptop has under 16 GB RAM. Track B if it has 16 GB or more.

Step 5 (Track A) · Cloud-free model, for under 16 GB RAM

Sign in to a free Ollama account (no card, no API key), then launch Claude Code on a low-usage cloud model:

ollama signin
ollama launch claude --model gemma4:31b-cloud
# fallback (also free, also low usage):
# ollama launch claude --model nemotron-3-nano:30b-cloud

Check your remaining free usage any time at ollama.com/settings. Reserve cloud usage for real file edits, plan and scaffold elsewhere (Section 06).

Step 5 (Track B) · Local model, for 16 GB+ RAM

Pull the largest model your RAM can hold (do this on good Wi-Fi at home, the download is several GB), then launch Claude Code on it. Local is unlimited, no quota:

ollama pull qwen3.5:9b          # 16 GB RAM. Use gemma4:26b or qwen3-coder-next if you have more.
ollama list                     # confirm the model is downloaded
ollama launch claude --model qwen3.5:9b

Step 6 · Test it (both tracks)

From any project folder, ask Claude Code to read your current folder or create a small file. If it runs the command instead of printing JSON, your stack works.

Prefer to wire it up by hand, or ollama launch is unavailable on your build? Point Claude Code at the Anthropic-compatible endpoint directly (works for a local model; for cloud, run ollama signin first):

export ANTHROPIC_BASE_URL=http://localhost:11434
export ANTHROPIC_AUTH_TOKEN=ollama
export ANTHROPIC_API_KEY=""
claude --model qwen3.5:9b

Environment variables tell Claude Code and Ollama how to behave. You only need a few. Set them before you launch, and open a fresh terminal after setting a permanent one or it will not apply.

Windows (PowerShell)

Permanent (recommended), then close and reopen the terminal:

setx OLLAMA_CONTEXT_LENGTH 65536

Just for the current window (temporary):

$env:OLLAMA_CONTEXT_LENGTH = "65536"

Check it worked: echo $env:OLLAMA_CONTEXT_LENGTH should print 65536.

macOS / Linux (bash or zsh)

For the current session:

export OLLAMA_CONTEXT_LENGTH=65536

To make it permanent, add that same line to ~/.zshrc (macOS) or ~/.bashrc (Linux), then run source ~/.zshrc.

Check it worked: echo $OLLAMA_CONTEXT_LENGTH should print 65536.

Work down this list in order. Most Sprint-morning problems are one of these.

Still stuck after the table? Email labs@nmgdigital.com with: your OS, ollama --version, claude --version, the exact command you ran and the full error text. Vague reports slow us down.

Hardware to bring on Sprint Day

• ☐Your own laptop with at least 8 GB RAM, a working keyboard and a working battery.
• ☐Your charger and power adapter.
• ☐A government photo ID (Aadhaar, College ID, Passport or Driving Licence).
• ☐Headphones (optional but recommended for the six-hour build).

Note: builders bring their own laptops. NMG provides a small number of backup laptops in genuine emergencies only.

Software installed and verified

• ☐Node.js v18.0.0+, verified with node --version.
• ☐Git, verified with git --version.
• ☐Ollama, verified with ollama --version.
• ☐Claude Code, verified with claude --version.
• ☐OLLAMA_CONTEXT_LENGTH=65536 set, terminal reopened, confirmed with echo (Section 03a).
• ☐Track A (under 16 GB RAM): signed in with ollama signin, and ollama launch claude --model gemma4:31b-cloud opens a working session that runs tools (not raw JSON).
• ☐Track B (16 GB+ RAM): model pulled at home (ollama list shows it), and ollama launch claude --model qwen3.5:9b opens a working session that runs tools.
• ☐Your editor of choice (VS Code recommended): code.visualstudio.com.

Accounts created

• ☐GitHub account with at least one public repo. Required for your Sprint Day submission.
• ☐Free Ollama account at ollama.com (run ollama signin). Needed for Track A cloud-free. No card.
• ☐Loom account (free) at loom.com. Required for your Sprint Day walkthrough video.

No Anthropic, Groq or other paid AI provider account is needed. Track A runs the model free on Ollama's servers, Track B runs it free on your own machine.

There is no cheat sheet and no day-of crash course. You are expected to self-study before Sprint Day until you can build with Claude Code and reason about SEO and digital marketing on your own. Work through every module below.

On Sprint morning we release the challenge brief and real, live client data at the venue. The exact task stays secret until then so nobody can pre-build a solution. It will be a real-world digital marketing automation. There is no cheat sheet, so arrive having studied Section 05.

Architecture prerequisites

To be eligible for scoring, your build must include:

• A central Claude Code Skill acting as the master coordinator.
• At least 2 sub-agents (for example one for extraction and one for formatting), triggered through Claude Code on Ollama.
• Use of Ollama web search to pull in live information during the build.
• A clean output pipeline that produces client-ready reports.

Required outputs

Your final submission must generate the same analysis in all three formats:

• A PDF report.
• An HTML report.
• A PPTX report.

Make your free compute last the whole six hours

If you are on Track A (cloud-free), the free tier is metered by compute time and it resets on a rolling window, so a careless builder can run out mid-build. Track B (local) has no quota, but a weak prompt still wastes time. Either way, work like this and you will finish:

• Plan and design off the clock model. Do your architecture, file structure, schema and copy in a normal free browser chat (claude.ai free, Gemini, ChatGPT) or on paper. Spend Claude Code only on actually writing and editing files.
• Scaffold before Sprint Day. Project setup, dependencies, empty components and config are allowed to be ready in advance. Then Sprint compute goes only to the new feature logic, not boilerplate.
• One precise instruction per call. "Write only the GA4 fetch function with this signature." Vague prompts make the agent loop and re-read files, which burns compute and time. This is the skill we are testing.
• Keep context tight. Work in a small focused folder. Run /clear between unrelated tasks. Do not hold a 50-file repo in context, it makes every call heavier.
• Use your own account. Each builder has their own free quota. Never share an account, a shared account drains instantly.
• Know your fallback ladder. Track A running low, switch to the other low-usage cloud model. Still stuck and you have 16 GB+ RAM, drop to a local model (unlimited). Worst case, finish the last edits by hand.

Check your remaining cloud usage any time at ollama.com/settings. Ollama emails you at 90%.

We enforce strict submission rules to keep things fair. Full legal and behavioural detail lives on our Terms and Conditions and Code of Conduct pages.

The 8 disqualification triggers

• Private repo: the GitHub repository is private at the deadline. It must be public.
• Missing Loom: the Loom video is missing, heavily edited or features another presenter.
• No logs: the unedited agent-log.md chat file is missing from the repository.
• No commit history: a single giant initial commit instead of clear, logical building phases.
• Auto-tester failure: the skill fails to run in the auto-tester within 5 minutes.
• Code reading score of 0: you cannot explain any block of your own code during demos.
• Plagiarism: 75% or more code duplication with any existing public repository.
• Conduct violation: breach of anti-harassment, safety or intellectual honesty rules at the venue.

Have questions? Read our Main Page FAQ or email us at labs@nmgdigital.com.