Skip to main content

The bots have arrived: run qbash from Muse and Grok Bot

Alex Blom

Agentic work decides its path during each run, and a deterministic Task follows a path written in Python and calls a model for bounded judgment

Your qbash Tasks, Memories, and connectors can now be called from Meta Muse and Grok Bot. This edition also covers an overhaul of MCP servers, v2 of our Task Builder, a laundry list of new models and connectors, three new posts, and the AI Ticker for September 25 to October 1.

Run qbash Tasks from Muse and Grok Bot

Muse, Grok, Dots: it seems the majors all believe that always-on personal bots, equal parts cute and intelligent, are the future (and I, for one, welcome them). Your Tasks, Memories, and third-party connectors in qbash can be called by your fuzzy new pal.

  • Meta Muse does not support custom MCP servers in its consumer app (Muse Code does), so it calls a Task's synchronous API through a Custom Connector.
  • Grok Bot connects to a qbash MCP server in the desktop app, and can use the Tasks, Memories, and connector tools you include.

Docs: Connect a qbash Task to Meta Muse · Connect qbash to Grok Bot

MCP servers now combine Tasks, Memories, and connectors

Add Task tools, Memory folder tools, and selected connector tools to the same server. Clients see only what you include, and connector credentials stay in qbash.

A support server, for example, can search your knowledge base, look up CRM records, and run a ticket-triage Task through the same URL.

Choose tools in bulk, mark a connector's tools read only, and write server instructions for clients to read before they call anything.

Read the MCP Servers docs →

Task Builder v2 imports automations from other platforms

Export an automation from n8n, Claude Code, OpenClaw, or Codex and import it. Task Builder writes a plan for what it can convert, builds the Task, and talks you through the rest.

  • More inputs. Attach PDFs, images, spreadsheets, and slides to a builder turn.
  • More to edit. Task Builder now creates and edits inline Markdown documents, scripts, and prompts, and covers triggers, Task linking, and secrets.
  • Test runs skip writes by default. Choose Test run (allow writes) when one run should write to connected systems.
  • A refreshed editor. Moving between code and builder is easier, and past builder sessions are saved. Each AI turn can be accepted, rejected, or reverted.

Open Task Builder →

GPT-6.1 Sol, Sonnet 5.5, and 30 OpenRouter models

  • GPT-6.1 Sol (gpt-6.1-sol) costs $2 / $10 per million input/output tokens with 1.05M context and reasoning effort from low to max. Above 272K input tokens, long-context rates apply.
  • Claude Sonnet 5.5 (claude-sonnet-5-5) keeps Sonnet 5's $2 / $10 price with 1M context. Adaptive thinking is on by default.
  • 30 OpenRouter models, including Kimi K3 and Z.ai GLM 5.3. OpenRouter models now support tool calling and the MCP proxy.
  • Updated Gemini and xAI models, including Gemini 3.8 Flash and Grok 4.7.

All of these are on by default with your own provider keys. Compare a new model with your current one in Playgrounds before switching a Task. Model Providers covers enabling models and adding keys.

Browse models →

Four new connectors

Browserless
Control a headless Chrome browser: inspect a page's interactive elements, run automation, and capture screenshots.
Stripe
Read revenue, billing, and payment data. Writes are limited to creating and updating customers and sending invoices.
Shopify
Read orders, customers, products, and inventory. Tag orders, customers, or products.
Notion
Query databases and data sources, read pages, and create pages.
Browse connectors →

Three new posts on bots and Jev

Grok Bot, Muse, and where thinking happens
Alex Blom
Agents suit unknown routes. Work that repeats is easier to trust as code.
I built a Jev + Claude loop to stop AI slop
Mitchell Tessier
Claude drafts and Jev grades, and failed drafts go back for another round.
Why I'm paying attention to Jev
Mitchell Tessier
Jev makes bounded decisions, and Mitchell thinks the narrowness is the point.

The AI Ticker, September 25 to October 1

Access got tighter and agents got harder to audit.

Some users report Sonnet 5.5's new cyber safeguards refusing legitimate security audits. One says blue-team work stopped mid-project. Anthropic's Cyber Verification Program does not yet cover 5.5 models.

Gemini 4 Argon tops the Vals Index, but access is limited to vetted security partners. Introductory pricing is $2 / $10 per million tokens, rising to $4 / $20 later.

OpenAI says research agents under reduced safeguards exposed 53 user images. During a cybersecurity evaluation, the agents posted them to image-hosting sites as unlisted links. OpenAI cannot identify the owners to notify them.

Researchers report that agents will alter their own audit logs. Direct requests, malicious skills, and reward incentives all worked in their tests. Arcjet now enforces Open Policy Agent rules before each agent action runs.

If you connect a qbash Task to Muse or Grok Bot, tell us what you had it do.

Video transcript

We're just about to get our first pictures from inside the spacecraft with average not Homer Simpson and we'd like to Ladies and gentlemen, we've just lost the picture but uh what we've seen speaks for itself. The Car of Air spacecraft has apparently been taken over conquered if you will by a master race of giant space ants. It's difficult to tell from this vantage point whether they will consume the captive earth man or merely enslave them. One thing is for certain, there is no stopping them. The ants will soon be here.

And I for one welcome our new insect overlords. Like to remind them that as a trusted TV personality I I can be helpful in rounding up others to toil in their underground sugar caves.

Stay in the loop

Product updates, tutorials, and AI insights. No spam.