glean
Your coding sessions are full of insights you're losing. glean catches them.
glean is a Claude Code plugin that quietly mines finished sessions for the moments that mattered - the surprises, the "huh, that didn't work", the times reality disagreed with your model. It writes them to disk as structured fragments you can review later.
This is reflective capture, not content extraction. Whether you're a developer who writes, or a developer who wants better recall of what you're learning, glean gives you a running log of the moments worth remembering.
Data flow
Claude Code session
|
v
SessionEnd hook (fires once, < 500ms, non-blocking)
|
v
Queue (.pending) ---------+
| |
v |
Worker picks up session |
| |
v |
Filter transcript |
Strip secrets (regex) |
Check dedup hash |
| |
v |
Haiku extraction via claude CLI
|
v
Parse fragments + concepts
|
v
Write to vault / markdown dir -> .delivered record
| (queue file removed)
failed write
|
v
Outbox (retry later)
Hooks stay dumb. All real work happens in the worker, in the background, after your session ends.
What it captures
glean is not summarizing your session. It is hunting for story fragments - specific moments worth remembering. Each fragment has five fields:
- Moment - What actually happened. Concrete, with real values and file names.
- Surprise - The expectation that got violated in one sentence.
- Tension - The unresolved question this raises. A real question, not rhetorical.
- Verbatim - A quotable line from the session, if there was one.
- Thread - A hook line for the post, essay, or followup this could become.
Most sessions produce zero fragments. That is correct. The prompt is tuned to say NO_IDEAS aggressively rather than force extraction.
Example fragment (truncated):
# The one knob that fixed everything
## What happened
Spent two hours chasing a stale bundle in the Remotion render. Tried clearing
node_modules, rebuilding the bundle, restarting Chrome. The fix was a single
cache-busting flag we'd disabled six months ago for "performance"...
## The surprise
The performance optimization from six months ago was actively making every
render silently wrong.
## The tension
How many of our other "performance" flags are active footguns?
## Thread
The cache flag we disabled for speed was silently breaking every render.
What leaves your machine
Be explicit: glean calls the Anthropic API (via the claude CLI using your existing OAuth). Here is exactly what is sent:
- Filtered transcript text. Only
human and assistant text messages. Tool calls, tool results, system prompts, and raw file blobs are stripped.
- Secrets stripped first via regex, before the API call. Patterns redacted:
- AWS keys (
AKIA[0-9A-Z]{16}, aws_secret_access_key)
- GitHub tokens (
ghp_, gho_, ghu_, ghs_, ghr_)
- OpenAI / Anthropic keys (
sk-..., sk-ant-...)
- Google API keys (
AIza[0-9A-Za-z\-_]{35})
- Slack tokens (
xox[baprs]-...)
- Stripe keys (
sk_live_, rk_live_)
- Private keys (
-----BEGIN (RSA|EC|OPENSSH|PGP) PRIVATE KEY-----)
- Bearer tokens,
Authorization: headers
- Common
.env-style assignments ([A-Z_]+_(KEY|TOKEN|SECRET|PASSWORD)=...)
- Nothing else. No file contents outside the transcript. No tool invocation results. No environment variables. No git state.
The output (fragments) lives on your disk. No telemetry. No phone-home. Code is stdlib-only Python and auditable.
Installation
# From inside Claude Code
claude plugin install glean
# Then run the post-install setup
/install
The /install step will:
- Verify Python 3.9+ and the
claude CLI are available
- Ask whether you want Obsidian output (one file per fragment, wikilinks) or flat markdown (directory of .md files)
- Ask for the vault path + folder, or the output directory
- Run a canary extraction against your most recent Claude Code session so you can see output immediately
