Who it is forA developer who starts a fresh coding agent session on a project and wants it to know what was done before.
The problemA new agent session spends dozens of shell commands re-reading files it already saw last week.
The outcomeThe agent begins with relevant prior work in its first prompt and can answer with one kb_search call instead of many grep and ls commands.
What you need first
- Node.js with npm, to run `npm i -g @arihantdeva/heimdall`
- macOS or a Debian, Ubuntu, Fedora or Amazon Linux machine, with the build prerequisites listed in the README for the graft backend
- git, because llama.cpp is cloned at first build
- A supported agent harness: pi, Claude Code, Codex, Cursor, Windsurf, or opencode
Steps
npm i -g @arihantdeva/heimdallheimdall init --harness claude-codeheimdall setupheimdall doctor- Open a new agent session in the project folder and ask it to call kb_search on <topic> before it starts editing
Prompt for your agent
In the ArihantDeva/heimdall setup, start a session in <project-folder> and call kb_search on <topic> before any shell exploration. Expected output: a short list of prior work with verdicts and paths. Verify that the first useful action uses a STRONG hit or that the agent says a hit is WEAK.
Caveat. The README names only some harnesses as wired by `heimdall init`, and the exact injection behavior for each harness is not described in detail, so test it in your own harness before relying on it.
Use this in your AI agent
Ask your agent
Using the Deduped MCP, get the setup steps for ArihantDeva/heimdall and "Start agent sessions with relevant prior work already loaded"
Add Deduped to Claude Code once
claude mcp add --transport http deduped https://dedupedai.com/mcp
Every Monday: the 5 hottest repos and exactly how to use them. Free, one email a week, unsubscribe anytime.
Get them freeShare: X LinkedIn Hacker News Email Copy link
Third-party repos we don't own. AI-written use cases can be wrong. Disclaimer