Build Your Own Wiki OR Clone anyapp with DeepSeek Harness for FREE
Install DeepSeek Harness, wire REA and a free OpenRouter preset, then prompt a local personal wiki.
Free models change names and rate limits. Treat every model slug below as “whatever the OpenRouter picker shows today,” and keep API keys out of git and chat logs.
What each piece does
- DeepSeek Harness is the agent that writes files, runs commands, and fixes its own errors. Official docs: deepseek.com/en/harness.
- REA (Reverse Engineer Anything) is a local MCP server. Harness calls it when a task needs to inspect files or software on your machine. Repo: github.com/morluto/rea. Inspect only software you own or are allowed to examine.
- OpenRouter is one API in front of many models, including free ones. A preset is a saved model list with fallback, so a rate limit does not stop the build.
- Free LLM Models NVIDIA NeMo-3 Ultra (free), NVIDIA NeMo-3.5 SuperVL (free), and NVIDIA NeMo-3.5 Lightning (free).
Prerequisites
- Node.js 20 or newer.
node -vshould printv20or higher. Node 24 LTS is a safe choice. Harness and REA both install throughnpx. - A Google account for DeepSeek Platform and OpenRouter.
- A terminal, and a folder where the wiki may be created (an empty directory is enough).
Step 1: Start DeepSeek Harness
The video uses Harness Desktop (drag the Mac app into Applications, or run the Windows installer). On Linux, and whenever you want the documented path, start the web UI from npm:
npx @deepseek-ai/dsh web
That serves the UI at http://127.0.0.1:3080 and opens a browser on a local launch. Pass --no-open if you only want the URL printed.
Leave this process running. Later steps assume Harness is open.
Step 2: Add a DeepSeek API key
Harness asks you to sign in or add an API key. Use the key:
- Open platform.deepseek.com and sign in with Google.
- Open API keys, create a key, name it (for example
harness-local), and copy it once. - Back in Harness, choose Add API key, paste the key, and continue.
In Settings, leave the agent preset on Standard. The video sets thinking effort to High on DeepSeek’s own models. You will switch the chat model to the OpenRouter preset before the wiki build, so this only matters for the first DeepSeek-hosted turns.
Step 3: Register REA with Harness
The client id for DeepSeek Harness is dsh. It needs rea-agents 6.5.0 or newer, which @latest selects.
npx rea-agents@latest setup --client dsh
npm may ask before it downloads the package. REA then prints a plan (the REA skill, Harness settings, and often Hopper, an analysis helper). Approve with y only after you have read the paths it will change.
Check the install:
npx rea-agents@latest doctor
You want a clean report. If a line is red, copy that output into a Harness chat and ask it to fix the reported check. Do not paste your API keys into that message.
Quit Harness completely and start it again so it loads the new MCP server. A running app can keep the old tool list.
Step 4: Connect OpenRouter
- Sign up at openrouter.ai (Google is fine) and choose an individual account.
- Create an API key and store it in a password manager. Skip adding a card if you will stay on free models. A key with no credit still fails paid models with HTTP 402, which looks like a bad URL if you are not expecting it.
- In Harness open Settings → Models → Add a custom provider.
- Name it
OpenRouter free. - Base URL:
https://openrouter.ai/api/v1 - API protocol:
openai-completions(OpenRouter speaks the OpenAI-compatible completions API). - Paste the OpenRouter key into the provider’s key field, or point the provider at an environment variable and export it before you launch Harness:
export OPENROUTER_API_KEY="sk-or-..."
npx @deepseek-ai/dsh web
The durable file is $DSH_HOME/settings.yaml (often ~/.dsh/settings.yaml). Prefer apiKeyEnv over pasting the secret into that file:
llm-pi-ai:
providers:
openrouter:
apiKeyEnv: OPENROUTER_API_KEY
api: openai-completions
baseURL: https://openrouter.ai/api/v1
models: []
You will fill models in the next step. The provider id stays lowercase. Renaming it later orphans saved sessions.
Step 5: Save a free-model preset
Free models hit limits mid-task. A preset lets OpenRouter try the next model on the list.
- In OpenRouter open Presets and create one. The video uses the name
free modelsand the descriptionfree AI model fallback for DeepSeek plus REA. - Add three free models. The November 2026 video picks NVIDIA NeMo-3 Ultra (free), NVIDIA NeMo-3.5 SuperVL (free), and NVIDIA NeMo-3.5 Lightning (free). If those rows are gone, add three current free models that support tools. Copy each slug from the model page; a wrong slug fails per request with 404.
- In the preset’s tools section, enable web search, date and time, and the sub-agent options the form shows (the video also turns on Fusion and Advisor).
- Save. OpenRouter shows a preset id shaped like
@preset/free-models. Copy it. - In Harness, under the OpenRouter provider, add a model. Model id is the preset id. Display name:
free models. Apply.
Select free models in a new chat before you build anything.
Step 6: Prove the tools answer
In an empty project folder, start a chat on the OpenRouter preset and send:
Check that the REA MCP server is connected. List the REA tools you can call, by name only.
Then read README.md in this workspace. If it is missing, say so and stop.
Do not inspect any app, binary, or folder outside this workspace.
Reply with: connected or not, the tool names, and a three-bullet summary of the file.
A good reply names tools and either summarizes your README or says the file is absent. If Harness claims REA is missing, rerun npx rea-agents@latest doctor, restart Harness, and try once more.
Step 7: Build the wiki
Still in that chat, with free models selected, send this prompt. It asks for your own product, not a copy of someone else’s name, colors, or assets.
Create a local personal wiki called Field Notes in ./field-notes.
It is my notes app. Do not copy any company's name, logo, colors, or layout.
Use an ink-and-paper look: off-white page, near-black text, one accent color #3d5a40.
Stack: a small Vite + vanilla JavaScript app. No paid APIs. No accounts.
npm install and npm run dev must work. Print the URL when you finish.
Pages, stored as markdown files under content/:
- home: what Field Notes is and how to add a page
- getting-started: the DeepSeek Harness, REA, and OpenRouter setup, in my own words
- projects: a short list of things I am building
- glossary: definitions for Harness, REA, OpenRouter, and MCP
UI:
- left sidebar with those four links
- a search box that filters sidebar titles as I type
- the selected page rendered from its markdown file
- a New page button that writes a new markdown file and shows it in the sidebar
- previous and next links between pages
Seed the four pages with real sentences, not lorem ipsum.
When the app runs, click through every sidebar link, search for "MCP",
create a page titled "Today", and fix anything that fails.
Harness should plan the folder, write the files, install dependencies, and start the dev server. Open the URL it prints. Check the four seeded pages, type MCP in search, and create Today. If a click does nothing, paste the browser error back into the same chat and ask it to fix that failure.
When the session stops because a free model is capped, stay on the same preset. Fallback is the point of the list. If every model is capped, wait for the limit to reset and tell Harness to continue in the existing field-notes folder instead of starting over.
What you have when it works
- Harness running locally, with a DeepSeek key for the product itself.
- REA registered for the
dshclient, anddoctorclean. - An OpenRouter provider whose only model is your free preset.
- A
field-notesapp you can edit by talking to the same chat.
The wiki is a local demo with files on disk. It is not hosted, not multi-user, and not a substitute for a real team wiki until you add auth, backups, and a place to deploy it.
Key takeaways
- Install order matters: Harness, then
npx rea-agents@latest setup --client dsh, then a full restart, then the OpenRouter provider, then the preset. - Put secrets in environment variables. A 402 means the key has no credit for a paid model. A 404 means the model slug is wrong.
- Free presets are a fallback list, not unlimited capacity.
- Point the first build at a product you own. Field Notes is enough to prove the loop: prompt, files, running UI, then a fix turn.
Subscribe & Follow
Get notified of new technical articles on AI/ML, Java, Python, and system architecture.