Ernest Pedapati, MD

As neurons are not meant to be alone,my posts about data science, research findings, and other musings are to connect to the larger neuroscientific community.

[Ernest Pedapati, MD @CBrainlab]

Say hello to sciClaw🦞🧪 - your paired scientist.
An open-source local research assistant, searches pubmed, writes and revises word docs with tracked changes, and logs everything for reproducibility. MIT.
Built by a physician-scientist on top of @SipeedIO's picoclaw, inspired by @steipete.
Custom new CLIs for pubmed, docx, audits, phi-cleaning, etc.
ChatGPT login OR bring your own AI model. YT tutorial, docs, github, preprint in thread 🧵
Contributors welcome!!!
#sciClaw #OpenClaw #agents #OpenSource #Science

[Ernest Pedapati, MD @CBrainlab]

sciClaw v0.3.2 is out

Addons are here optional sidecars for webtop browser desktops and Jupyter notebooks.

Also: gpt-5.5 and tighter security

brew upgrade sciclaw

https://t.co/C1hVa2Ap4J

[Ernest Pedapati, MD @CBrainlab]

Hotfix: sciClaw v0.2.8 — OpenAI changed how Codex streams responses and every message went blank. Found it, fixed it, shipped it same day. If your bot was saying "did not produce a user-facing reply" — upgrade now.

brew upgrade sciclaw && sciclaw service refresh

[Ernest Pedapati, MD @CBrainlab]

sciClaw 101 - starting a project with a collaborator

In Discord:
1. co-scientist accepts invite and joins server
2. setup a private channel for the project
3. add the co-scientist and sciclaw users

In sciClaw admin panel
1. grant user permissions to sciclaw bot
2. route new channel to a shared/cloud folder
3. share folder with co-scientist

tag sciclaw and get to work!

[Ernest Pedapati, MD @CBrainlab]

sciClaw is a research agent built for how labs actually work. Multiple scientists share one AI through Discord. Drop a dataset, ask for an analysis, and your collaborator picks up right where you left off.
Built for bench-to-manuscript:
- PubMed search and citation graphs
- Word documents with tracked changes
- Figure generation
- Full evidence trail your PI can review

We just launched our community Discord.
Come talk about your workflows, get help setting up, or shape what we build next. → https://t.co/IBFAzG9SMe

[Ernest Pedapati, MD @CBrainlab]

@steipete setting the pace, time to upgrade sciClaw to also have MCP capabilities. Brilliant. https://t.co/2XjuoeeRrM

[Ernest Pedapati, MD @CBrainlab]

Why this matters: sciClaw agents run multi-hour research sessions in Discord — literature searches, document drafting, data analysis. Without context compaction, the agent either hits the window limit and loses memory, or you pay for 80% stale history on every turn.

ctxclaw gives us a clean boundary: the runtime does agent work, ctxclaw decides what old context stays live. sciClaw v0.2.6 ships with ctxclaw integration built in.

brew install drpedapati/tap/ctxclaw

[Ernest Pedapati, MD @CBrainlab]

We kept hitting the same problem running sciClaw in long Discord sessions: the agent burns most of its context window replaying stale tool outputs nobody needs anymore. A 2000-line file read from 40 turns ago is still sitting in history, and every new turn pays for it.

[Ernest Pedapati, MD @CBrainlab]

ctxclaw sits between your agent runtime and the model call. It reads the assembled prompt envelope from stdin, identifies bloat, and emits a smaller prompt on stdout.

What it actually does:
- moves oversized old tool outputs to archive files
- replaces stale conversation history with a checkpoint summary
- preserves the recent, structurally valid tail

It's not an agent. Not a memory store. Just a prompt-context optimizer.

https://t.co/Tcv0tlfHIK

[Ernest Pedapati, MD @CBrainlab]

New CLI: `sciclaw prompt inspect --session `. Shows exactly how many tokens each part of your prompt uses — identity, bootstrap files, skills, memory, history, tool schemas. Debug context window issues without guessing.

[Ernest Pedapati, MD @CBrainlab]

sciClaw v0.2.6 is out — the first stable release with a full browser-based web UI. Manage channels, routing, models, skills, and a new System tab that diffs your workspace personality files against the release templates.

brew update && brew upgrade sciclaw
sciclaw service refresh

https://t.co/mppROpsoKG

[Ernest Pedapati, MD @CBrainlab]

Adding a Discord channel to sciClaw routing used to mean copying a channel ID from Discord settings. Now the routing wizard auto-discovers your servers and channels — pick from a searchable list. Also added a folder browser for workspace paths. No more typing paths blind.

[Ernest Pedapati, MD @CBrainlab]

I went back to chatgpt 5.4 to make a formatted reference list. I even gave it the actual pubmed ids. 2/4 were still hallucinated.

Save yourself the anxiety and make sure the agent uses a proper tool to directly query the database. Try my MIT pubmed-cli tool for zero hallucination free reference searches. https://t.co/tpfTwlwTgn
used in https://t.co/4OHl47xOHL too!

[Ernest Pedapati, MD @CBrainlab]

Try asking your agent working on one repo to investigate a bug on dependency repo and log a pull request. Lots of wins where your whole ecosystem gets stronger.

[Ernest Pedapati, MD @CBrainlab]

Discord use is cleaner now.

This release adds:
- clearer job/status behavior
- real slash-command support for /btw and /skill
- better support for Anthropic OAuth users through the Claude bridge

If you already use sciClaw, upgrading is straightforward:
https://t.co/S1gEruHnLr

[Ernest Pedapati, MD @CBrainlab]

sciClaw v0.2.4 is out.
This release is mostly about making sciClaw easier to use in real research work, especially in shared chat channels and longer scientific workflows.

brew tap drpedapati/tap
brew install sciclaw

to upgrade:
brew upgrade sciclaw

Release notes: https://t.co/nTTCZgKFqp

[Ernest Pedapati, MD @CBrainlab]

Research workflows got stronger too.

This release improved built-in support for:
- PubMed and citation lookup
- document-based workflows
- weather and other simple utility lookups
- more reliable handling of uploaded files and follow-up tasks

So sciClaw is better at staying grounded in the actual materials you’re working with.