Claude Code
Run inside the take-home repository to add the server at local scope.
claude mcp add --scope local inside-the-build -- npx -y inside-the-build@beta
OPEN-SOURCE BETA · LOCAL-FIRST · TELEMETRY-FREE
A take-home shows what they built. See how they built it.
Inside the Build turns Claude Code or Codex sessions from an engineering take-home into an inspectable process brief: decisions, corrections, verification steps, and questions worth exploring in the next conversation.
No scoring. No server or hidden upload. An inspectable file stays local until shared.
Install
Add one local MCP server
Requires Node.js 20+, a Git repository, and macOS or Linux.
Only Claude Code and Codex are supported in the v0.1 beta.
Codex
Run inside the take-home repository, then confirm the MCP server is available.
codex mcp add inside-the-build -- npx -y inside-the-build@beta
After the assignment, start a fresh agent session and invoke
submit_takehome first. It writes
<repository>-inside-the-build.md. Inspect that file before sharing it;
prompts can contain secrets or personal information and the beta has no built-in redaction.
The problem
Polished code is only the endpoint
A repository can show whether the solution works. It rarely shows how the engineer framed the problem, responded when an approach failed, directed their tools, or verified the result.
Inside the Build adds that missing context without introducing another test.
The submitted code shows
- The final architecture
- Implementation quality
- Test coverage
- Scope and completeness
Inside the Build shows
- How the candidate approached the problem
- Decisions and tradeoffs they encountered
- Corrections and changes in direction
- How they worked with AI coding tools
- Specific areas to discuss in the interview
What you receive
A process brief, not a scorecard
Inside the Build asks the local writing agent for a descriptive draft. It does not rank, grade, or make a hiring recommendation.
Synthetic example · No candidate data
Example API service — Inside the Build
- The candidate began by asking the agent to inspect the service before proposing changes.
- The first implementation put validation in the route, then moved it to middleware.
- A failing integration test exposed stale fixture data rather than an API regression.
Example follow-up question
What made middleware a better boundary for validation than the original route-level approach?
[4] (claude-code, synthetic-01): “Move the validation into middleware; both handlers will need the same rules.”
How it works
Add it to the take-home you already use
The candidate completes the assignment
They work normally with Claude Code or Codex. No recording software or special development environment is required.
They generate their report locally
After finishing, the candidate runs Inside the Build from the assignment repository.
Inside the Build reconstructs the process
It finds matching sessions and asks the local agent for a descriptive draft with quoted receipts. Production does not yet verify those citations automatically.
The candidate submits the report
The Markdown file can be inspected and shared through the organization’s existing take-home workflow. Inside the Build never receives it.
Trust
Designed to inform a conversation—not surveil a candidate
Inspectable artifact
The output is an ordinary Markdown file that can be opened before it is shared.
Local by design
No Inside the Build server, database, account, hidden upload, or runtime telemetry. Processing happens on the same machine as the coding agent.
Bounded to the project
Event extraction is limited to sessions whose recorded working directory belongs to the current Git repository or one of its worktrees.
Descriptive, never evaluative
The report is explicitly prohibited from scoring the candidate or making claims about their competence.
Open source
Hiring teams, candidates, and contributors can inspect what is read, excluded, and produced under the Apache-2.0 license.
Limitations
Beta evidence has boundaries
Inside the Build provides a model-written draft from recorded work in Claude Code or Codex. It does not prove authorship, detect cheating, measure competence, or replace a technical conversation.
- Cross-session chronology is not guaranteed.
- Malformed session lines may be omitted without an exact count.
- Citation verification is not connected to production generation.
- Prompt injection, model variance, and missing sessions can affect the draft.
- Prompts can contain secrets; the beta has no built-in redaction.
Read the full known limitations before using it in a real workflow.
Make the process part of the work sample
Add Inside the Build to your next engineering take-home and understand more than the final diff.