17 tools, one pip install, for agents whose web fetch hits anti-bot walls.
Scrapling's MCP server offers fingerprint-impersonating HTTP requests or a real stealth browser for pages that block a plain fetch.
Install: pip install "scrapling[ai]" then scrapling install
Claude Code: claude mcp add ScraplingServer "/path/to/scrapling-mcp"
Before v0.4.13, use scrapling mcp.
Save cookie/session state across runs, take screenshots, and inspect network request history for debugging.
For anyone whose agent needs to read docs, pricing or research pages and keeps hitting a block.
https://t.co/jEgLdOGGSo
https://t.co/c2l1i4nHci
One slow MCP task used to hold your turn open. Now it moves to the background.
Claude Code immediately backgrounds an MCP tool call that runs as a task, then tells Claude the result when it finishes.
Run `claude --version`: needs 2.1.319 or newer. Current stable is 2.1.322 as of 2026-11-12.
CLAUDE_CODE_DISABLE_MCP_TASK_BACKGROUND forces the old blocking wait; CLAUDE_CODE_DISABLE_MCP_TASKS=1 turns MCP background tasks off entirely.
Useful if your MCP servers run slow tools: search, a scrape, or a long build.
https://t.co/HMrLHU9bWd
Your coding agent can create a DynamoDB table, write an item and read it back on 127.0.0.1:9297 instead of your AWS account. No AWS account needed, no service fee.
That's AWS Emulator, which AWS wrote up on Nov 11. Before, the agent's AWS tests meant deploy waits, real resources and real keys in the shell. Now the same CLI and SDK calls get answered on your laptop.
Linux: install the RPM or Debian package, then run aws-emulator. macOS and Windows: AWS publishes no official prebuilt image. Build the image from the Dockerfile in the docs, create ~/.aws/emulator, and start the container with docker run, publishing 127.0.0.1:9297 and mounting that directory. Copy the exact command from the Docker docs, because the volume path differs on Windows. The binary runs inside the container, not on the host. Once the endpoint is listening, in the agent's shell:
export AWS_ENDPOINT_URL=https://127.0.0.1:9297
export AWS_SHARED_CREDENTIALS_FILE=~/.aws/emulator/9297/emulator_credentials
export AWS_CONFIG_FILE=~/.aws/emulator/9297/config
export AWS_REGION=us-east-1
unset AWS_ACCESS_KEY_ID AWS_SECRET_ACCESS_KEY AWS_SESSION_TOKEN AWS_PROFILE
The AWS post sets a few more variables; copy the full list from there. Then aws sts get-caller-identity should show account 000000000000.
Limits: it doesn't evaluate IAM policies the way AWS does, so don't use it for security tests. One account, no cross-Region behavior. Operations it doesn't support yet return HTTP 501. Check the per-service fidelity scores in the docs first.
https://t.co/JWUGjs1xo4
From Dec 8, the start date in GitHub's Nov 10 changelog, Dependabot alerts turn off on most repos with no activity in the past year. The same archival behavior has been on the public roadmap since Nov 5 (issue 1350). That can include the side project that still runs but hasn't had a push in a year.
Activity means a push to any branch or tag, creating the repo, or turning on Dependabot alerts. Existing alerts are archived, not deleted.
After that, new alerts for those repos stop arriving. Any push turns them back on, or you can turn them on in the repo's settings. Either way, the archived alerts come back and the one-year clock starts over.
Rough list of your source (non-fork) repos with no push since 2025-12-08 (pushedAt only sees pushes, so treat it as a starting point):
gh repo list --limit 1000 --no-archived --source --json nameWithOwner,pushedAt --jq '.[] | select(.pushedAt < "2025-12-08") | .nameWithOwner'
Any repo on that list that still runs in production: push to it before Dec 8, or turn its alerts back on in settings after.
https://t.co/XoLi7IhkZT
A high-severity GitHub advisory for mcp-server-git was published Nov 11 at 14:25 Vietnam time, with the fix in pull request 5514 and on PyPI as 2026.11.11. The issue had been reported publicly on Nov 8 (issue 5488): a read-only git_status on an untrusted repo could run a command from its own .git/config, such as its fsmonitor setting, with the server's privileges, in versions through 2026.11.4.
Other settings could run programs too: an external diff tool, hooks, filter and merge drivers, signing.
The fix overrides those settings on every Git command and empties repo-defined filter/merge drivers. Global drivers like Git LFS keep working. Git older than 2.31 is refused.
A plain git clone does not carry it. A project archive, CI artifact, shared folder, vendored tree or copied agent workspace that keeps .git can.
Check which mcp-server-git version you run.
Upgrade to 2026.11.11. You need Git 2.31+.
Until upgraded, point the server only at repos whose .git directory you trust.
https://t.co/qePhYdWPx8
https://t.co/ZGG2thIn8E
Gemini CLI's /restore could delete or overwrite files you excluded with .git/info/exclude instead of .gitignore.
Only with checkpointing on: excluded files created after a checkpoint got deleted, and excluded files present before it got their recent changes overwritten with old content. Same for a global excludes file.
3 selected fixes in Gemini CLI v0.68.0:
- /restore now honors those excludes, as it already did .gitignore.
- A manual gemini-3.1-pro-preview selection could silently revert to gemini-3-flash-preview on later turns, including after cancel + /rewind, while the footer and /model still showed gemini-3.1-pro-preview. The fix preserves the selection.
- 429/quota retries now show a cool-down countdown on attempt 1 instead of looking hung.
Check: gemini --version
If it is below 0.68.0, update.
Until updated, keep local-only files in .gitignore if you rely on /restore.
https://t.co/h6eiQeCSD1
https://t.co/6fmmQ9W0tR
Claude Code passed 57% of the web task checks without guides and 88% with them. That's from the project's own evals (Nov 8 snapshot, 215 tasks, Opus 5.5). Codex with GPT-6.1 Sol: 54% to 82%.
The gap: coding agents often default to older JavaScript patterns from their training data, even when the browser has native HTML or CSS for the job, like animating dialogs and popovers.
Modern Web Guidance reached 1.0 on Nov 10, out of preview. It's an open-source skill plus a command-line tool that puts targeted guides into your agent's context. Code is Apache 2.0, search runs offline with no API key, and it's now stewarded by the Chrome, Edge, Firefox and MDN teams.
Try it on an existing repo first:
npx modern-web-guidance@latest legacy-scan .
The full Claude Code and Codex plugin setup steps are in the README.
The scanner uses static rules that cover roughly 20% of guides so far. Treat its list as a starting point, then have the agent modernize one finding at a time.
https://t.co/d07rvk22jn
Once a month: sign in to each supplier portal, download the invoices, file them for your accountant. n8n's free, open-source agent Nod is built to take that job, working in its own sandboxed browser.
It's a research preview from n8n Labs, a separate app for Mac and Windows under Apache 2.0. No n8n account needed. You bring the AI: an API key, your ChatGPT plan, or a local model.
Before you hand it a card, check the money controls. Every task shows its cost and ends with a receipt. You can set a monthly spending limit. It checks in with you once a task passes $5, and you can change that amount. Some early testers found tasks cost more than they expected.
Before it sends, pays, posts or deletes anything you didn't ask for, a separate check confirms you asked. n8n says that check is itself a model and can make mistakes.
Logins and cards stay encrypted on your machine, entered through forms the model can't read. It stops when your laptop sleeps.
Start with a read-and-download job. Set the spending limit before you connect a card.
https://t.co/ky7clxlUTD
One sentence in the prompt moved Claude Fable 5.2 from merging repeated code in under a tenth of tries to about half, in Anthropicโs testing on small coding tasks.
The sentence:
"This codebase is yours: what it's for and how it's built are your call, and you'll keep building on it."
Rates varied by task. The requested feature worked in nearly every try either way.
Cost: about 1.5 times the output tokens, plus unrequested changes. Apart from bug fixes, existing feature behavior changed in one try in six to ten, and more often at higher effort.
If you want the model to improve the structure as it works, consider adding it. If you want the code changed as little as possible, such as a surgical patch, leave it out. A "No drive-by refactors, renames or cleanups" rule earlier in the prompt doesnโt cancel it: merges still happened in about a third of tries.
Read the diff: closing summaries often missed small changes.
https://t.co/02uYEbcX4p
3 selected permission and secret fixes in Claude Code 2.1.321.
npm latest is 2.1.322; npm stable is still 2.1.311 (checked today).
The release notes say it fixed:
โข MCP OAuth requests sending secrets such as refresh tokens to another host when a server redirected them.
โข A case where a read-only Bash command ran outside the sandbox without a permission prompt.
โข Deny rules and protected-file prompts not applying to the file that a cp or mv into a directory writes.
Native installs default to the latest channel. The claude-code Homebrew cask and the stable apt, dnf and apk repos follow stable, which is typically about a week old and skips releases with major regressions.
Run claude --version. If it's below 2.1.321 and you rely on these protections: on a native install, make sure "autoUpdatesChannel" isn't set to "stable", then run claude update. Homebrew does not auto-update. For the latest cask, run brew upgrade claude-code@latest. Or wait for stable.
https://t.co/vSg9hSq9VA
https://t.co/ttMJGEXcDK
No origin content-signal header? Converted pages get:
Content-Signal: ai-train=yes, search=yes, ai-input=yes
To change that, set your own content-signal header at the origin. Cloudflare keeps your value.
Markdown for Agents left beta on Nov 9 with dashboard analytics, logs and a playground.
Dashboard: zone > Agent Readiness > Markdown for Agents.
Analytics shows top user agents, top converted pages, failed/partial pages and estimated token reduction. Logs has sampled events. Playground compares Markdown with HTML for a path.
Conversion: Pro, Business, Enterprise and SSL for SaaS only. User agents are self-reported strings, not verified bots. Token savings use bytes / 4 estimates, not tokenizer counts. HTML input is capped at 6 MiB after decompression.
With conversion enabled, test your own page:
curl -H "Accept: text/markdown" YOUR_PAGE_URL
https://t.co/mqju9yvBIv
https://t.co/8f2c2AStH0
Claude says โall tests pass.โ The last test ran before its last edit.
AgentMon Live adds a done gate before Claude stops: it checks recorded activity for stale checks, runs with zero tests, and open todos. Gaps go back to Claude once per prompt: finish the work or say plainly why not.
Free, Apache 2.0, by Codenotary; released Nov 9. Needs Claude Code 2.1.308+.
/plugin install agentmon-live --marketplace codenotary/agentmon-live
The web view has no login. On a shared machine, set web to off.
The optional second-model check sends ~12,000 characters of session text per call, at most 30 calls per session by default. Set advisorMode to gate to turn model calls off.
Use /agentmon check to inspect what's unfinished without sending it back to Claude.
https://t.co/e7uuO9t45N
https://t.co/M08FsdQvKK
Your coding agent can run on a schedule in your repo folder and leave the result in a chat you check from your phone.
Open WebUI added coding agents in 0.13.1, then agent automations in 0.13.3 on Nov 8.
Instead of opening a terminal for each recurring chore, you can schedule it and read the result later.
Set the terminal, working folder, agent (such as Codex or Claude Code), and schedule. Runs reuse that terminal and folder; results land in a chat.
Needs an Open Terminal that offers agents; Tether connects your own machine. The terminal and agent must still be available when the run starts.
Go straight to 0.13.3; 0.13.2 fixed plain-HTTP chats. Back up the database before upgrading: schema changes are included. For multi-worker, multi-server or load-balanced setups, update all instances at once; rolling updates are unsupported and will cause app failures from incompatible schemas.
One untested schedule idea: audit dependencies in one repo folder and leave a short summary in the chat.
https://t.co/LG9drdhbXx
https://t.co/lR84UCamnt
3 prompts: a 14 KB C extension vs ~300 KB in Rust, then one push made C about 2x faster.
While walking the dog, Simon Willison asked Claude chat (Opus 5.5) to build parse_number(), a small SQLite function, in C and Rust and compare speed, code complexity and file size.
First: C was smaller, but slower on mixed data: 67 ns/row vs Rust's 41-46.
Second: write a C decimal parser that catches up to or beats Rust. A ~30-line fast path reached 16-19 ns/row across the three tested datasets, bit-for-bit identical across 8M fuzz checks.
Third: turn the work into a research report in his repo.
Rust was still shorter and safer to write; C needed manual buffers and overflow checks.
One small function, 2 million rows, a 2 vCPU Linux VM. The report is AI-generated.
On a small function like this, have the agent build both, check correctness on representative inputs, and measure speed, size and lines. Give the slower one a specific target, then save a searchable write-up.
https://t.co/KfO3pHcayo
Claude Code opened an unmentioned skill 0 times in 30 turns. Naming it: rule followed in 30 of 30.
Garry Tan's open-source harness evals, published this weekend, tested a "check gbrain first" rule on Opus 5.5. gbrain is his own tool.
The skill's name appeared in every request, but it stayed unread. Adding "Use the brain-first skill." to the prompt got the rule followed in 30 of 30 turns, vs 18 for the unmentioned skill and 17 for no rule.
The same rule in the system prompt or startup file reached 30 of 30 turns too.
Small dev test: 10 public tasks, one repetition, pinned versions, one model per harness. These Claude Code results are descriptive: Opus already checked gbrain in 17-18 of 30 turns without a rule.
For a workflow that depends on a skill, its name belongs in the prompt.
https://t.co/zHRfvv6lmO
A scripted agent run can end before its background tests report back.
Before: opencode run exited when the model's first reply ended, cutting off test suites and evals before the model saw their results.
After: in OpenCode 2.2.6, it waits for background shell commands, subagents and image jobs its prompt started, then lets the model respond to their results before exiting.
While it waits, it prints "waiting for N background job(s)". With --format json, you get a background_wait event.
There is no time limit. A dev server or file watcher that never exits keeps the run open until Ctrl+C or SIGTERM. Set a timeout on CI jobs. SIGTERM cancels the session and background jobs, exiting 143.
Update to 2.2.6 or later.
https://t.co/dDeeGkK7hK
https://t.co/w9eqyHRYaq
A permission prompt can leave your agent waiting while you're away.
Before: a tool call waits for you to answer.
After: the call is denied and the turn continues.
Claude Code 2.1.319 adds this setting. In your Claude Code settings, set:
{
"permissions": {
"denyInsteadOfPrompting": true
}
}
This covers terminal sessions, subagents and teammates.
The call that would prompt gets denied. It does not get auto-approved.
Run claude --version first. You need 2.1.319+.
https://t.co/yDoJxuMoGo
https://t.co/1y5CAk8WNR
Running several coding agents still leaves you creating and organizing threads yourself.
For a solo builder, that's another project to manage.
Amp Spaces puts that work in one conversation with a coordinator agent. It spawns agents and organizes their threads. You get a bird's-eye view and can still dive into the details when necessary.
Amp's security-review example split a codebase into 16 areas, reviewed each in its own thread, checked every finding in another, and kept a summary artifact.
To start: click + in the Spaces sidebar, set a title and project, then tell the agent the goal.
https://t.co/BLjnw8zrwN
Finished agent work sits unread in another session.
Several Codex agents are running. You hunt through panes to find which one finished.
Open the agent command center with:
`codex agents --no-daemon`
The command center and session picker now mark unread results. You can also switch between connected backends from the terminal session picker and keep your drafts.
Run `codex --version` to check: you need 0.167.0+.
https://t.co/0Stjan1s95
Your agent agreed to โdo this tomorrow.โ That often means a database, a cron job polling for due work, and a redeploy when timing changes.
Use Vercel Schedules for the handoff. Define a static schedule in vercel.json, or create one at runtime with the Schedules SDK.
Static schedules can target a private function, an HTTP path, or a Queue topic (a named inbox for work). Runtime schedules always target a Queue topic.
Add a schedule/v1beta trigger to make a function private, with no public URL. Enable, disable, or run schedules on demand without redeploying. Static definitions still sync on deploy.
Hobby includes 500 schedule creations and 5,000 runs per month. Existing Cron Jobs still run.
https://t.co/9hz3Ehlf6Q
https://t.co/7u3vckejPk