Learn while your agent works — and carry project context forward.
Named for Ekalavya, who mastered archery practicing before a silent statue of his guru. Here, the statue talks back.
Learn while your agent works — local knowledge graph, spaced repetition, and quiz gating for agent-assisted development.
2026-11-14
Sterne
—
Tools
—
Uptime
—
Eklavya is a plugin for Claude Code. It does two jobs:
It helps you understand the code Claude writes. It asks you short
questions about the work, while the work happens.
It remembers what happened in your project. The next time you open
Claude, it can recall what you did last time.
Your learning history and memory are stored on your computer. There is no
Eklavya account. Eklavya sends anonymous daily usage counts (including which dashboard
sections and explainers are used), never code, paths, names or text. Every field is listed,
and eklavya telemetry off stops them. New installs use claude-sonnet-5
through your Claude Code login for memory summaries and prompt feedback, so
captured session evidence goes to that model. Memory and data
explains the controls.
Eklavya notices the ideas behind Claude's work, like "database index" or
"refresh token". Then it asks you about them.
What it does
How it works
Read more
Asks while you work
One question at a time, in the middle of the task (never in the first five minutes of a session), written so you can answer without having read the code, and so the right option is not the longest, the most detailed or explained in advance, and does not always come first. Answer, say you don't know and get the explanation, or skip and move on; a question that was itself at fault never counts against you, and an automatic question is meant to stay one question.
For a larger change, Eklavya asks Claude to hand the building to a background agent (in your task's worktree, one stage at a time for a staged plan) and ask you questions while it works. If Claude starts changing a second file itself, Eklavya reminds it once.
Eklavya records what you and Claude do. It turns that into short notes called
observations. When you start a new session, Claude gets recent, relevant
notes to check against the code.
What it does
How it works
Read more
Captures your work
Saves prompts, file edits and tool calls as you work; unattended claude -p runs keep only tool and file activity.
Gives Claude past work and where the last session left off at session start, on related prompts (not on short ones with fewer than three meaningful words, unless an entry names them), and before it reads a file with history. On prompts and on a resume it skips entries the current session wrote, which Claude already has. The session banner shows the estimated context saved.
Memory is only evidence. It never marks a topic as learned. Only a real answer
does that.
Artifacts
Ask Claude to explain something "as an Eklavya artifact". You get an HTML page
with colour-coded diagrams, chosen to fit the idea (sequence, swimlane, state
machine, bars and more), that you can print or save as PDF.
What it does
How it works
Read more
Explainer pages
"Explain refresh-token rotation as an Eklavya artifact."
On by default: after a wrong answer, a page restates the question, gives a short answer, and explains the idea with a diagram, in the background. Then correct your answer from it in the dashboard, and the miss counts as right in your accuracy and toward your level.
Pages are saved on your computer, listed in the dashboard and opened there as tabs; a page made from the Eklavya template scrolls with the dashboard as one page, and its HTML button saves the page itself, free of the viewer's script.
Run eklavya dashboard to open a local page in your browser. It keeps running
in the background, and each session start restarts it if needed. The sidebar's
foot shows the version it is running, with links to the docs. A malformed
request gets an error reply instead of stopping it. A page left open shows
new activity as it happens, including a session summary rewritten in place,
and never moves what you are reading. You can also
install it as an app
and open it from Spotlight, the Dock or the Start menu. On your first visits,
tips point out features;
dismiss them one by one or switch them off under Settings, then Companion. The archer companion walks toward you with a random verb (Pondering…) while loading and appears beside empty states and feedback; the same tab previews its 24 expressions and its loading walk and has the daily card switch. On a wide window
each page sits centred beside the sidebar. It has five parts:
Part
What you can see
Read more
Learning
Accuracy, mastery, missed answers still to be corrected, your streak and its calendar, with the archer beside it and one card the first time you open the page each day, reviews due with a command per project to start them, the top three shown (each run asks the weakest first), weak topics, sessions and projects, each project at its own level.
Reviews one of your prompts using what later prompts had to add, and offers a better version. On for new installs; existing installs keep their choices. Needs memory and an observer model. The next review waits until you acknowledge or delete the waiting one; the session greeting flags it in amber. It never changes your scores.
For a short summary in chat, use /eklavya:progress.
The dials
Each setting works on its own. Memory has its own switch, so you can turn
questions off and keep memory on.
Dial
Default
Other choices
quiz.enabled
true: ask questions
false: no questions
quiz.enforced
false: commit freely
true: pass a quiz before a commit
quiz.panel
true: experimental side panel on supported Claude Code builds
false: Claude's question card. Cowork and unsupported hosts use the card.
focus
concept: the idea behind the code
project: this codebase · learn: a topic you pick
feedback.enabled
true: review one of your prompts using what you had to add later
false: no prompt review. Needs memory and an observer model.
cadence
as-you-go: during the task
end: after the task
difficulty
auto: grows with you
easy, medium or hard
memory.enabled
true: record and recall
false: learning only
By default Eklavya asks only in sessions that change code (edit tools or
shell commands, in this checkout, a sibling worktree or a repository below the
folder you started in; a git pull or branch switch alone does not count), up to four
questions per session on its own. Checkpoint questions are at least four minutes
apart; questions while background agents build share the budget but skip that
gap. /eklavya:quiz can always ask more. Change a dial with /eklavya:mode in Claude Code,
eklavya config set in a terminal (eklavya config get quiz.enabled reads one back), or the dashboard's Settings pages. quiz.panel can be changed with eklavya config set quiz.panel false or in Settings; a round you ask for with /eklavya:quiz keeps its topic and length across the panel's Next button, and a miss in the panel starts its explainer page in the background (and so does its orange New to me — Teach me button, which records "I don't know" even with explain_on_wrong off), asking once more and queuing a prompt for Claude only if Claude Code refuses or fails that start twice. A picked answer gets bold text, a thicker border and a filled background; on the desktop app each answer is a numbered card you can click anywhere on, Teach me is a compact button with its whole body clickable, the Other box is a three-line text area, and the header shows the mascot as vector art (the terminal draws the same head in coloured block characters). In the panel's Other box, Enter keeps what you typed and moves the keyboard to Submit answer. A prompt the panel sends Claude shows in your transcript as one short line. /eklavya-panel-demo opens a mock question to try the panel's clicks and a small picture test; nothing it does is recorded. If Claude Code's auto mode denies the panel's tools, allow them as the manual shows, or ask Claude to. See the dials
and every setting.
Commands
Command
What it does
/eklavya:quiz [topic]
Take a quiz now.
/eklavya:mode [value]
Show or change the dials.
/eklavya:skip
Skip questions for this session only.
/eklavya:learn <topic>
Take a short lesson.
/eklavya:level [level]
See your level or pick one.
/eklavya:pack [topic]
Add new topics, or topics from your code.
/eklavya:progress
See what you know and what is due.
/eklavya:memory
Search past work or save a note.
/eklavya:gate
See why a commit is blocked.
/eklavya:setup
Check the installation without changing settings.
You can also just ask, like "Space out Eklavya's questions". See
all commands and
the CLI.
Install
With Claude Code and Node.js 22+ with npm on macOS, Linux or Windows:
npm install -g eklavya && eklavya install
The installer uses the defaults without asking settings questions and keeps
any choices you already saved. Allow Eklavya's tools in auto mode,
then restart Claude Code and start a task. Eklavya updates itself in the
background. See Installation options
for other setups.
Development setup, the manual test scripts and the release process are in
CONTRIBUTING.md. CI fails any change that leaves a line,
branch or function of the runtime untested (npm run coverage in mcp/).
Noch keine Tool-Liste. Remote-Server erhalten eine nach dem ersten Health-Check.
FAQ
Fragen
Wie installiere ich Eklavya?
Kopieren Sie auf dieser Seite die stdio-Konfiguration für Cursor, Claude Desktop, Claude Code oder VS Code. Sie nutzt die Installation des Herausgebers, keinen PlugMCP-Proxy.
Benötigt Eklavya einen API-Schlüssel?
In diesem Eintrag ist keine erforderliche Umgebungsvariable aufgeführt.
Welchen Transport verwendet Eklavya?
Eklavya läuft lokal über stdio.
Wann wurde Eklavya zuletzt geprüft?
Zuletzt geprüft: 2026-11-14. Learn while your agent works — local knowledge graph, spaced repetition, and quiz gating for agent-assisted development.
