Skip to main content
This guide walks you through creating an experiment, connecting Claude Code via MCP, and running an implementation. By the end you’ll have a tracked experiment with a linked PR.

​
1. Set Up Your Profile

Your profile helps Remyx tailor recommendations to your work.
  1. Click Account (bottom-left of the sidebar)
  2. Set your Role (ML Engineer, Researcher, Data Scientist, Product Leader, etc.)
  3. Add a one-line Interests description
  4. Click Save
You can skip this and come back later under Account > Profile Details.

​
2. Create a Research Interest

Research Interests tell Remyx what to track across papers, repos, models, and more. Navigate to Feed in the sidebar and click + New Interest. Click Refresh to get your first batch of recommendations immediately.

​
3. Create Your First Experiment

Navigate to Outcomes in the sidebar. Click + New Experiment.
  1. Set the Source type to Paper
  2. Search for a paper relevant to your target repo
  3. Fill in:
    • Name: Something descriptive (e.g., “Multi-hop retrieval for complex queries”)
    • Hypothesis: What you expect (e.g., “Adding multi-hop retrieval will improve resolution rate for multi-topic queries”)
    • Target metric: e.g., resolution_rate
    • Tags: e.g., retrieval, rag, multi-hop
    • Target repo: owner/repo (e.g., remyxai/VQASynth)
    • Project: Select or type one (e.g., Customer Support AI)
  4. Click Create
You land on the experiment detail page. The Origin section loads the paper title and links immediately. Within a few seconds, the launch context appears: key method badges, abstract excerpt, Docker image, target repo, and an implementation plan grounded in your repo’s actual file paths.

​
4. Install Claude Code and Connect Remyx

​
Install Claude Code

Verify the installation:
If claude is not found, open a new terminal window (PATH needs to refresh).

​
Authenticate

This opens your browser. Log in with your Anthropic account (Pro/Max/Team/Enterprise). Follow the prompts to authorize. This is a one-time setup. Type /exit to quit when done.

​
Add Remyx as an MCP Server

Replace YOUR_REMYX_API_TOKEN with a personal (rmxu_) key created under Account > API Access in Remyx. See API Scopes for how keys and scopes work.

​
Verify the Connection

You should see:
You can also test from inside a Claude Code session:
Then type:
Claude Code should call the Remyx MCP tools and return your experiments.

​
5. Run the Implementation

  1. cd into your target repo
  2. Start Claude Code:
  3. Tell it to implement the experiment:
Claude Code fetches the experiment context via MCP, reads your repo files, generates the implementation, creates a branch, pushes it, and opens a PR. The PR is linked back to the experiment automatically. Watch for permission prompts (git push, etc.) and approve them.

​
6. Check the Result

Refresh the experiment detail page in Remyx. You should see:
  • A green PR banner at the top with the PR title, status, and GitHub link
  • The PR listed in the Resources sidebar
  • Events in the Activity feed (PR opened, implementation summary)
Click the PR link to review the code changes on GitHub.

​
7. Log a Decision

Once the experiment has results, log what the team decided and why. This is the most important step: it captures the reasoning that would otherwise live in someone’s head. Scroll to the Decision section and click to write:
“Positive result, +3.1% on resolution rate. The re-ranker helps with multi-topic tickets where the old retriever returned tangentially related articles. Ship to production.”
This decision is timestamped, attributed, and visible to the whole team. When someone new joins and asks “why do we use multi-hop retrieval?”, the answer is here.

​
8. Check Insights

After creating several experiments with shared tags, navigate to Insights in the sidebar. Remyx groups experiments by tag, computes hit rates per direction, and shows which directions produce consistent results. It also recommends next experiments based on your team’s history.

​
Next Steps

Outcomes

Full guide to the experiment timeline and detail views

Connectors

Connect GitHub, Linear, Jira, and Slack

ExperimentOps Concepts

The methodology behind systematic experimentation
See the Feed docs for how resource discovery and Research Interests work in detail.