YOUR AGENT IS BURNING TOKENS BEFORE IT DOES A SINGLE F*CKING THING In my Grok Bot article, the expensive part isn’t just the answer. It’s every browser step and every routine that wakes up to find nothing new 5 repos I’d use to cut that waste in an agent stack I control: MCP Servers ▸ reference Fetch, Git and Filesystem tools that show how to give an agent direct access without browser clicks. This isn’t a ready-made Gmail connector. ▸ github.com/modelcontextproto… MCP Compressor ▸ shows the agent a small tool list, then loads the full schema only for the tool it picks ▸ github.com/atlassian-labs/mc… Token Optimizer MCP ▸ measures and reduces context returned to local coding agents ▸ github.com/ooples/token-opti… Langfuse ▸ traces calls, tokens and cost so you can find the expensive step ▸ github.com/langfuse/langfuse Trigger dev ▸ runs scheduled jobs with logs and retry controls, outside an endless agent chat ▸ github.com/triggerdotdev/tri… → first remove work the agent never needed to do. A routine checking every 15 minutes still wakes up 96 times a day, even if its reply is one line These tools won’t change Grok Bot’s quota for you. Save them for workflows where you control the calls.

HOLY SH*T, AN APPROVE BUTTON IS USELESS IF THE AGENT MADE UP THE INVOICE AMOUNT My Grok Bot vs Dots article asks one question: would you let an agent find an overdue invoice and press Send? For a self-hosted version, these 4 repos address different parts of that decision: Invariant ▸ can block a get_inbox → send_email tool-call sequence when the recipient is outside your company ▸ github.com/invariantlabs-ai/… Deliberate ▸ pauses a LangGraph agent and shows the approver the amount, reasoning and evidence you pass in ▸ github.com/beomwookang/delib… Paperless-ngx ▸ gives the agent a searchable invoice archive to cite instead of guessing from memory ▸ github.com/paperless-ngx/pap… Bubblewrap ▸ limits which files a local Linux agent process can see ▸ github.com/containers/bubble… → these aren’t add-ons for Grok Bot or Dots. They show what I’d want to control in my own agent. And a tool-call rule won’t catch a browser click unless that action passes through the same guard Before I approve Send, show me the invoice, the amount and the recipient. Save this for that test

THE TEST SAID “EXACT MATCH.” THE CODE WAS F*CKING WRONG In one agent-driven game reconstruction, the new code used 0.0 where the original used 128.0. The comparison still passed. Once the checker was fixed, an audit found 12 more errors in functions it had already accepted That’s the part of reverse engineering with agents I can’t stop thinking about 3 repos behind the story: REA ▸ lets a coding agent investigate an app through MCP or CLI and return findings with evidence ▸ github.com/morluto/rea Ghidra ▸ the native-binary analysis engine REA can use to inspect functions, calls and assembly ▸ github.com/NationalSecurityA… PhotoCraft ▸ a Rust photo editor built from public specs and observed behaviour, without copying proprietary code ▸ github.com/storytold/photocr… → REA and Ghidra help you understand software you’re allowed to inspect. PhotoCraft shows what an independent build can look like. But a green test means little until you’ve given that test a wrong answer and watched it fail Save this before trusting an agent’s “exact match.”

WHY THE F*CK DOES A BOT GET YOUR INBOX AND THE SEND BUTTON AT THE SAME TIME? In my article, I compare Grok Bot and Dots on one task: find an overdue invoice, check the amount, and draft an email. These 4 repos would let me test that workflow in my own setup: Cua ▸ gives the agent a separate computer instead of a shared desktop ▸ github.com/trycua/cua Playwright MCP ▸ can run in an isolated browser session, away from your logged-in mail tabs ▸ github.com/microsoft/playwri… OpenBao ▸ stores credentials so your app can fetch them for a task without passing them to the model ▸ github.com/openbao/openbao Phoenix ▸ can trace the invoice source, the amount used, and any send call you instrument ▸ github.com/Arize-ai/phoenix → none of these tools blocks Send by itself. I’d start with 10 test invoices: the agent shows the source, checks the amount, and prepares a draft. Sending stays disabled. If I can’t trace where a number came from, the agent isn’t getting that button. Save this before connecting an agent to your work inbox

WHY THE F*CK IS YOUR AGENT WRITING AN ESSAY TO ROUTE ONE SUPPORT TICKET? Jev takes context and returns a choice, score, or probability. Its weights are closed, but these 4 repos let you build and test the same kind of decision layer: TypeSafe Skills ▸ teaches your coding agent how to frame questions for Jev ▸ github.com/typesafe-ai/skill… TypeSafe Python SDK ▸ calls Jev from your code and returns structured decisions ▸ github.com/typesafe-ai/types… Strands Decider ▸ an open, local 2B model for choices, scores, and yes/no decisions ▸ github.com/strands-labs/stra… Laya ▸ another local engine for the same types of decisions ▸ github.com/NandhaKishorM/lay… → I’d run the same 100 labeled tickets through all three models and check where each sends a double-charge complaint. TypeSafe’s speed benchmarks won’t tell you the error rate on your data. High confidence won’t, either. Save this before you put a decision model in charge of your support queue.

IF YOUR AGENT USES CLAUDE TO SORT EVERY SUPPORT EMAIL, YOU’RE PAYING FRONTIER PRICES TO PICK A F*CKING FOLDER. the Jev article makes a useful distinction: the model that writes a reply doesn’t have to be the model that decides which queue gets the email 5 repos to test that idea on your own tickets: TypeSafe Skills the official skill teaches your coding agent to build with Jev’s Choice, Score, and Noul questions. ▸ github.com/typesafe-ai/skill… System One Adapter send the same typed questions through OpenAI, Anthropic, or Gemini APIs. compare their routes with Jev on labeled tickets instead of trusting a cost or speed demo. ▸ github.com/typesafe-ai/syste… Strands Decider a local 2B decision model for picking options and scoring them with confidence. one way to test the idea without a TypeSafe API key. ▸ github.com/strands-labs/stra… Laya another local option: Choice, Score, and Noul, plus a router that selects a checkpoint for the request. its checkpoints download on first use. ▸ github.com/NandhaKishorM/lay… jev-starter the layer after the prediction: thresholds, fallbacks, evals, and human review. a confidence of 0.85 is a setting to test, not an accuracy guarantee. ▸ github.com/hamakyo/jev-start… → let the cheap layer sort the queue. call the expensive model when an email actually needs a written reply. keep Send behind a separate decision you control. Jev itself isn’t open-source; the local models here are alternatives, not its released weights. save this before you route another inbox with a frontier model.

DID YOU BLOCK THE GMAIL PLUGIN? THE AGENT CAN STILL ACCESS YOUR EMAIL THROUGH THE BROWSER, DAMN IT If your email is already open in the browser, blocking the connector won’t close it. This is one of the pitfalls I discussed in my article about Grok Bot and Dots I found 4 repositories for those who want to build their own agent with clearer access boundaries: 1. Steel Browser Don’t want to run the agent in your personal Chrome? Give it a separate, managed session. The key is not to transfer your profile or cookies into it. ▸ github.com/steel-dev/steel-b… 2. Sandbox Runtime Want to limit which files and domains the agent process can access? Rules are set at the OS level, without requiring Docker. Read access must be configured separately: by default, the entire disk is not hidden. ▸ github.com/anthropics/sandbo… 3. Cerbos Need to prevent emails from being sent until the recipient and amount are verified? You can define a rule for the “send” action in the policy. However, Cerbos itself doesn’t see the button and won’t prompt a user: verification and confirmation must be built into the workflow before the action takes place. ▸ github.com/cerbos/cerbos 4. Daytona Don’t want to store files from all tasks on a single machine? Create a separate sandbox for each job and close it afterward. This is an environment for code and tools, not a ready-made browser agent. ▸ github.com/daytonaio/daytona → Four bots with different roles but sharing a single email session still have shared access. I would first separate the browser and stop Send, and only then assign roles to the agents these are tools for your own workflow, not settings that you can simply enable in Grok Bot or Dots save your changes before you give the agent access to your work email

A SINGLE AGENT CAN ALREADY CREATE ANOTHER GROK BOT FOR YOU. WHO THE HELL IS MONITORING ACCESS RIGHTS? In the article, I compared Grok Bot and Dots using a task involving email and a “Send” button. And now, not only have collections of plugins appeared around them, but also tools for managing the bots themselves Here are 5 repositories that show where things are headed: 1. awesome-grok-bot-plugins Want to quickly browse plugins? Here are 219 entries from a Cursor Marketplace snapshot as of August 12. This is a catalog for a specific date; it’s no guarantee that every item is available right now. ▸ github.com/rdmgator12/awesom… 2. awesome-dots Guides, tools, and scripts for Dots all in one place. This is an unofficial list of resources, not a set of ready-made agents. ▸ github.com/mergisi/awesome-d… 3. grokbot-fleet An MCP for managing a list of bots and updating their profiles, routines, and skills. There’s no official Grok Bot management API, so the project uses a bridge via the bot’s computer. It’s still in its early stages—I wouldn’t give it access to production accounts without testing it first. ▸ github.com/Dahhrk/grokbot-fl… 4. grok-bot-skill Allows Cursor, Claude Code, or Codex to write a Grok Bot and create a new one directly from the terminal. Requires macOS with the Grok Bot app installed and an active session. ▸ github.com/adamanz/grok-bot-… 5. awesome-grok-bots A small collection of profiles tailored to specific tasks, skills, and plugins. This is useful if you want to start with a bot that drafts a message rather than sending it right away. ▸ github.com/doanbactam/awesom… → It’s now easier to create another bot. However, bots belonging to the same user (Grok Bot) still share the same cloud computer and its sessions. A new role in the profile does not create a new access boundary. Save your changes before linking your work email to this command

GROK BOT GIVES YOU A COMMAND. SHE ONLY HAS ONE F*CKING COMPUTER In the article, I discussed the moment when the agent found an overdue invoice, drafted a letter, and got as far as “Send.” Now I’ve found four repositories for those who want to build their own version of this assistant or at least figure out what permissions it grants 1. OpenDots Want your own workspace with permanent agents? Here, you can configure a separate computer, MCP tools, and confirmation prompts before actions for each Dot. This is an early prototype, not a ready-made replacement for Dots: conversations and computers require separate service configuration. ▸ github.com/CopilotKit/OpenDo… 2. awesome-grok-bot Already using Grok Bot? This directory lists 60+ skills, plugins, MCP servers, and alternatives. It’s a useful resource before connecting another tool to your work account. This is an unofficial directory. ▸ github.com/ZeroPointRepo/awe… 3. awesome-openai-dots Want to figure out what tasks to assign to Dots? This collection includes applications, rules, and scenarios with human-verified dots. Some of the scenarios were suggested by the community and aren’t ready to use out of the box. ▸ github.com/VoltAgent/awesome… 4. awesome-grokbot Want to quickly assign Sales, Eng, and Ops roles? Here you’ll find profiles and templates for building a team of bots. Just remember: multiple Grok Bots belonging to the same user still share the same cloud computer and sessions. ▸ github.com/mergisi/awesome-g… → The four repositories serve different purposes. The team template does not provide isolation, and the list of plugins does not determine who is authorized to click “Send.” Save your changes before granting the agent access to the client's email

WHAT THE F**K THE AGENT WAS GIVEN ACCESS TO EMAIL. BUT WHO WILL STOP IT BEFORE IT CLICKS SEND? In the article, I compared Grok Bot and Dots on a simple task: finding an overdue invoice, verifying the amount, and drafting an email. The most important question arises when the agent gains access to the browser and work accounts Here are 10 repositories I’d use to break down that access layer by layer. These aren’t 10 ready-made replacements for Grok Bot and Dots MY WORKSPACE 1. open-dots — an early self-hosted prototype with action confirmations; not a one-to-one replacement ▸ github.com/Anil-matcha/open-… 2. Firecracker — a separate microVM for the agent to run in ▸ github.com/firecracker-micro… 3. E2B — an environment where the agent runs code and tools separately from your computer ▸ github.com/e2b-dev/E2B 4. gVisor — an additional boundary between the agent container and the host ▸ github.com/google/gvisor BEFORE CLICKING THE SEND BUTTON 5. HumanLayer — human confirmation of actions; the authors warn that the public code is outdated ▸ github.com/humanlayer/humanl… 6. AgentGate — an experimental MCP gate: allow, prompt a human, or block ▸ github.com/CodedVibesX/agent… 7. OPA — rules for send, pay, and delete as code. Validation must be built in before the agent performs the action ▸ github.com/open-policy-agent… 8. Infisical — storing email and CRM keys outside of agent files ▸ github.com/Infisical/infisic… BROWSER AND TRACES 9. browser-use — browser automation. If you provide it with an authorized profile, the agent 10. Langfuse — tracks the agent’s actions if logging is enabled. The contents of emails and invoices in the logs will also need to be protected ▸ github.com/langfuse/langfuse → Blocking the email connector isn’t enough if the agent can still open email in a browser Save this list before you give the agent a production account and the permission to send emails

THIS IS F**CKING GOLD A CODING AGENT NEEDS MORE THAN JUST SKILLS I’m adding 10 repositories to the article about Codex and Claude Code. They don’t teach the agent to write code faster; instead, they help restrict access, verify results, and catch problems before deployment ACCESS BOUNDARIES 1. gVisor — isolates the agent container from the host kernel ▸ github.com/google/gvisor 2. Cerbos — access rules for actions: e.g., read-only access, but no deletion in production ▸ github.com/cerbos/cerbos 3. OpenBao — secret storage and management instead of keeping keys in the agent's folder ▸ github.com/openbao/openbao 4. detect-secrets — catches keys in the code before commit ▸ github.com/Yelp/detect-secre… COSTS AND ERRORS 5. LiteLLM — a unified model gateway with cost tracking and limits ▸ github.com/BerriAI/litellm 6. Phoenix — traces and evals to see where the agent went wrong. Important note: uses the ELv2 license, so it’s not classic open source ▸ github.com/Arize-ai/phoenix BEFORE ACTION AND MERGE 7. HumanLayer — a project focused on coding agents working with a human-in-the-loop. However, the authors already consider the code in this repository outdated; I wouldn't implement it as a ready-made approval layer right now ▸ github.com/humanlayer/humanl… 8. devcontainers CLI — a reproducible environment for the agent and checks ▸ github.com/devcontainers/cli 9. MSW — mocks APIs during tests while the agent verifies the UI. ▸ github.com/mswjs/msw 10. PR-Agent — reviews PRs before a human clicks merge; the repository has moved to a new GitHub account. ▸ github.com/The-PR-Agent/pr-a… Save this collection before granting the agent access to your working repository

HOLY SH*T, AN AGENT ISN'T ENOUGH TO AUTOMATE A BUSINESS I’m adding another 10 repositories to my previous list of 30. It’s not just about AI: you need documentation, a place to store leads, an email system, and a way to figure out where things go wrong. BUILDING THE ASSISTANT 1. Flowise — a visual agent builder. However, the repository is archived and the project has reached EOL; I wouldn't choose it for a new project ▸ github.com/FlowiseAI/Flowise 2. Langflow — visual AI workflows and components. ▸ github.com/langflow-ai/langf… 3. AnythingLLM — an assistant for working with your own documents. ▸ github.com/Mintplex-Labs/any… FEEDING IT DATA 4. Qdrant — vector search for a knowledge base. ▸ github.com/qdrant/qdrant 5. Docling — converts PDFs and other documents into a usable structure for further processing. ▸ github.com/docling-project/d… KEEPING THE CLIENT 6. NocoDB — spreadsheets for leads, statuses, and manual review. ▸ github.com/nocodb/nocodb 7. EspoCRM — a CRM for when spreadsheets just don't cut it anymore. ▸ github.com/espocrm/espocrm 8. listmonk — mailing lists and email campaigns. I’d still leave the actual sending to a human, though. ▸ github.com/listmonk/listmonk RUN AND TEST 9. Windmill — scripts, jobs, and workflows instead of yet another manual task. ▸ github.com/windmill-labs/win… 10. Langfuse — tracing AI steps: see exactly what the agent did and where it went wrong. ▸ github.com/langfuse/langfuse Save this list—especially if you already have an agent but your leads are still ending up in a random spreadsheet.

HOLY SH*T, JARVIS NEEDS MORE THAN JUST ONE PROMPT 20 Repositories for Your Personal AI Jarvis Using Claude Code In my article, I discussed how to assign real tasks to an assistant without handing over your entire business. Here are the tools for memory, voice, automation, documents, and access control CORE 1 Claude Code ▸ github.com/anthropics/claude… 2 MCP Servers ▸ github.com/modelcontextproto… 3 Agent Skills ▸ github.com/anthropics/skills MEMORY 4 claude-mem ▸ github.com/thedotmack/claude… 5 Mem0 ▸ github.com/mem0ai/mem0 6 Graphiti ▸ github.com/getzep/graphiti VOICE 7 Whisper ▸ github.com/openai/whisper 8 faster-whisper ▸ github.com/SYSTRAN/faster-wh… 9 Piper ▸ github.com/OHF-Voice/piper1-… TASKS AND CLIENTS 10 n8n ▸ github.com/n8n-io/n8n 11 Trigger.dev ▸ github.com/triggerdotdev/tri… 12 Windmill ▸ github.com/windmill-labs/win… 13 NocoDB ▸ github.com/nocodb/nocodb 14 Twenty ▸ github.com/twentyhq/twenty DOCUMENTS 15 Gotenberg ▸ github.com/gotenberg/gotenbe… 16 Stirling-PDF ▸ github.com/Stirling-Tools/St… SECRETS AND CONTROL 17 Infisical ▸ github.com/Infisical/infisic… 18 SOPS ▸ github.com/getsops/sops 19 Sandbox Runtime ▸ github.com/anthropics/sandbo… 20 Langfuse ▸ github.com/langfuse/langfuse Save this list if you're building something. an assistant for real-world work tasks ↓

HOLY SH*T, 15 MORE REPOSITORIES FOR MY ARTICLE ON CODEX AND CLAUDE CODE The first five are alternative coding agents. The rest help with project context, environment setup, testing, and releases AGENTS 1 Cline ▸ github.com/cline/cline 2 Roo Code ▸ github.com/RooCodeInc/Roo-Co… 3 Gemini CLI ▸ github.com/google-gemini/gem… 4 Crush ▸ github.com/charmbracelet/cru… 5 SWE-agent ▸ github.com/SWE-agent/SWE-age… CONTEXT AND BOUNDARIES 6 Serena ▸ github.com/oraios/serena 7 Daytona ▸ github.com/daytonaio/daytona 8 SOPS ▸ github.com/getsops/sops TESTING AND RELEASE 9 git-cliff ▸ github.com/orhun/git-cliff 10 Renovate ▸ github.com/renovatebot/renov… 11 uv ▸ github.com/astral-sh/uv 12 ShellCheck ▸ github.com/koalaman/shellche… 13 Difftastic ▸ github.com/Wilfred/difftasti… FILES AND MEMORY 14 MarkItDown ▸ github.com/microsoft/markitd… 15 Mem0 ▸ github.com/mem0ai/mem0 Save this collection for your next project ↓