Open-Weight AI Agents: Melaya's October 2026 Recap

Your model can work out what needs doing. October was about giving it somewhere to do it: your browser, your Android phone, your desktop apps and your business systems, with the tools and controls staying in place when you change the AI.

If you use OpenRouter, Qwen, Ollama, LM Studio or another open-weight model, you probably know the gap. A good answer is not the same as an updated customer record, a finished spreadsheet or a task completed in an app with no API.

Melaya connects those two worlds. You bring the AI. Melaya supplies the connected tools, device execution, knowledge, workflows and permission controls around it.

October 2026 brought Computer Control, faster browser and phone execution, Obsidian vaults, assistants in Slack and Microsoft Teams, standalone product subscriptions and a much deeper Odoo integration. This recap explains what changed, what it is useful for and where to start.

Open-weight AI agents: what October changed for your work

Here is the month in practical terms. These are additions and improvements to Melaya, not seven separate AI subscriptions you need to buy.

OpenRouter and Qwen tools: keep the model, add execution

An AI router helps you choose and reach models. A model generates an answer or proposes a tool call. Something still has to run that call, check access and return the result.

That is where Melaya fits. It is useful whether your reason for choosing an open-weight model is cost, deployment flexibility, customization or keeping particular workloads on your own infrastructure.

October made that setup easier with Sign in with OpenRouter, clearer recovery when a provider reaches its limit, and better detection of models running on your computer. The provider catalog also expanded with Eden AI and Sarvam AI. A provider connection and a model's actual tool or image capabilities still matter: a text-only model does not become a vision model because it has tools.

For MCP, the important detail is the client. An OpenRouter API key or a Qwen model alone does not connect to an MCP server. Use an MCP-capable application that supports your chosen model, or choose the model inside Melaya. Private OpenAI-compatible endpoints, including suitable vLLM or llama.cpp deployments, need the corresponding supported configuration and network access.

Read the model-router and open-weight setup guide or the Melaya MCP Server walkthrough. Your AI provider bills its inference separately; Melaya access does not include an unlimited supply of model tokens.

AI browser and Android automation: fewer trips back to the model

Imagine giving someone directions. You could phone them after every turn. Or you could say, “Open the page, enter these details, check the result, then come back if something looks wrong.”

October added the second approach through browser_fast and phone_fast, then improved it across the Assistant, pipelines and MCP. Your main model plans the work. Melaya carries out the short sequence against the live interface and returns control when it needs help.

Laya, Melaya's System One decision model, handles narrow judgments where needed, such as matching a description to a control or deciding whether a list row meets a condition. An exact visible label can be resolved without asking another model at all. Laya does not replace your main model or secretly plan an entire job.

One retained browser example makes the difference tangible: opening a Luma event, checking its panel, closing it and checking the list took one Fast Mode call and 6.1 seconds in a live Brave rerun on October 17. The October 6 one-step-at-a-time baseline took six calls and 108 seconds. These are dated observations of one task, not a controlled comparison of every model, site or connection.

On Android, a separate retained test collected 20 Reddit posts, including subreddit, age, votes and comments, in 16.2 seconds. That is a phone result, not a browser benchmark. We keep the surfaces separate because page loading, app behavior and the requested task change the result.

The more general improvement is fewer large-model round trips for routine work. It can reduce repeated context and waiting, without promising a fixed token saving on every job. See the System One explanation and Fast Mode product guide.

Browser Control also gained separate working tabs for up to eight concurrent agents, subject to your plan and available sessions. A tab held by one agent is not silently taken over by another. Background navigation became less disruptive, while the extension continues to enforce site access and the chosen autonomy mode.

On phones, you can target a particular device for a pipeline. Batch results better distinguish a locked phone, a missing accessibility service, a refused app and an action that did not complete. Those details save time when the problem is setup, not the model.

AI computer use: bring desktop apps into the workflow

Browser automation stops at the browser. October's new Melaya Computer Control extends the same idea to native desktop applications.

The desktop app brings the local runner into an installable application, with connected machines, model setup, update status and a visible way to stop control. Computer tools are available through the Assistant, pipelines and MCP when the relevant access is granted.

Computer Control can read application controls, operate forms and tables, and use screenshots or bounded visual targets when an interface does not expose enough structure. It also supports batches, stored plans and Fast Mode. A result that cannot be verified is reported as such rather than treated as a confirmed success.

For users, the key feature is taking over. Moving the mouse or using the keyboard can pause execution, release held input and let the system reassess before continuing. Stop remains available. On the desktop, approval notices show the action that needs attention rather than a generic “something happened” notification.

For teams, machine selection means a pipeline can stay on the computer you chose. If that machine is offline, Melaya reports the problem instead of quietly moving work elsewhere. Linux server and VM setups are also documented; platform capabilities and setup requirements differ.

Explore AI computer use with Melaya and the current desktop downloads. The Microsoft Store announcement followed on November 3; it is a later milestone, covered in the Windows launch article, rather than an October store-release claim.

AI assistant for Slack and Teams: work where the conversation is

October brought Melaya Assistant into Slack and Microsoft Teams. Mention it or send a direct message, use your linked Melaya account, choose a model and work with your connected services. Relevant thread context can accompany the request.

The useful part is what happens after the answer. Generated files can return to the conversation. Charts can be delivered as images. A person can review an approval in the supported chat flow, and actions that need the full app have a route back to Melaya. Slack also gained an App Home view with setup and connection status.

The Assistant itself became more capable at operating your Melaya workspace. It can discover the relevant workspace tools, create or inspect pipelines and help manage runs. Tools load when needed rather than forcing the entire catalog into every request. Large results and long context can be read in pages, making room for the actual task.

Each conversation now keeps its model, connectors, autonomy and devices. Reopening a chat should mean returning to that chat's setup. Stream recovery, clearer provider-limit messages and saved download cards also make longer work less fragile.

Follow the Slack integration guide or Microsoft Teams guide. Installation, workspace permissions and member linking still apply; a channel mention does not grant everybody access to somebody else's accounts.

Obsidian AI agents: use your vault without flattening it

Your notes are more than a folder of text. Links, headings, properties, tasks and daily notes carry context that a plain file upload can lose.

October introduced Obsidian vaults as a knowledge source across pipelines, Assistant conversations, the extension and MCP. You can work with a vault on your computer, an uploaded snapshot or a supported Git-backed vault. An optional cloud copy makes a local vault available when configured, while showing its age if the computer is offline.

For a local Qwen workflow, this gives you a practical starting point: keep a vault local, choose a suitable local model, and ask for a cited summary before permitting any changes. “Local model” alone is not a promise that the whole workflow stays local. Your vault settings, other connectors, execution target and any cloud services determine the actual data path.

Some Obsidian app operations require Obsidian itself and its supported CLI or Local REST API connection. Uploaded snapshots are not a running Obsidian app, and a supported subset of features is not every plugin you have installed.

AI agent builder: change models without rebuilding the process

The value of a visual agent builder grows when a workflow survives the next model release.

October added bulk AI provider and model changes in Monitoring, with a reviewable selection and undo. You can inspect the models used by individual agents and choose the runner machine for work that needs your computer. Device kits make it easier to add phone, browser or computer capabilities to the right agent.

The AI-assisted builder now retrieves relevant Melaya guidance while planning. A reasoning-only agent with an explicitly empty tool list stays tool-free. Quality observation became the default for pipelines, helping surface issues without automatically rewriting the workflow.

Teams gained a Runner role between viewing and editing. A teammate can run the pipelines they are allowed to use without receiving the right to redesign them. Team AI lets an owner configure the project's model access; project provider credentials are not handed to a member's own computer.

Scheduling also became more practical. A local run can be held while its runner is offline and started within the configured reconnect window. Owners receive notices for held, missed or paused schedules. Run visibility and reconnection improved, so a dropped live connection is less likely to make working pipelines appear stuck.

For an enterprise using open-weight models, the benefit is architectural: evaluate a new model while keeping the workflow's tools, access rules and review points. Model behavior can still change, so test the workflow after a switch. Reusable infrastructure removes repeated setup; it does not remove evaluation.

AI SEO and marketing automation: a fix list backed by your site

A model can sound confident about your SEO without having crawled your website. October's whole-site audit gives the conversation actual evidence.

Melaya Marketing crawls the site within the plan's limits, checks technical issues and joins available Search Console and Bing data. It examines internal links, redirects, canonicals, language tags, duplicate content, sitemap problems and search performance. Search-demand discovery adds real suggested searches, without inventing search volumes.

The report connects issues to pages and products. It can suggest which page to improve, which topic needs coverage and which technical fix to do first. The Site Explorer adds filters, page detail, changes between runs and exports. A disconnected browser can reattach to a running audit rather than starting the work again.

The audit also gained checks relevant to AI search visibility: crawler access, structured data, JavaScript-only shells and inconsistent bot policies. These identify obstacles to discovery; they do not guarantee an AI citation or a Google ranking.

Backlink workflows became more efficient at filling forms, reusing scrubbed form knowledge and remembering failures. A submission is distinguished from a verified live listing. Social workflows gained clearer unanswered-message signals, contact lists, per-platform routes through API tools, browser or phone, and review cards that show the actual item being answered.

Google Ads tools expanded to cover campaign and country reporting, PMax placements, conversion goals, Search setup, budgets and bids, network settings, search intent and asset diagnostics. Platform reporting limits still apply, and changing an ad account remains an explicit action. The point is to make a recommendation inspectable and executable, not to promise profitable campaigns.

Explore Melaya Marketing. The free audit summary and paid AI-report capabilities differ; use the pricing page for the current product and plan limits.

Odoo AI agents and business integrations: connect work to records

On October 26, Melaya became an approved Odoo Learning Partner. This is the Learning Partner relationship, not a Ready, Silver or Gold designation.

The Odoo connector moved to direct Odoo 20 JSON-2 operations, with typed actions and version-specific corrections across business workflows. That lets an agent perform a specific supported operation rather than treating an ERP as a vague chat interface.

The Odoo app integration adds a record panel, pipeline runs tied to records, a recent-runs view, approval follow-up and an Add from templates workflow. Templates explain which Odoo apps and other services they need. Operations use the connected project's access, with writes reviewed through the Odoo integration's approval rules.

See the Odoo partnership and Odoo setup guide for the requirements. An implemented integration is not a claim that every marketplace listing or certification is approved.

Other ecosystem changes widened the kinds of work you can connect:

The Tickstem launch article explains its role in more detail. New connector tools do not bypass vendor subscriptions, approval requirements or access scopes.

AI agent pricing: start with the product you need

October introduced a simpler way to buy Melaya: choose individual products, combine them for a bundle saving, or choose a full-platform plan when you need broader workflow and team capacity.

For example, Mobile Agents plus Browser Agents launched at $16.99/month, instead of $19.98 separately. The paid allowance applies to the selected product at its Bastion limits. It does not turn every other product or your whole account into Bastion, including when you use MCP.

The final October setup also keeps the Sandbox allowance alongside standalone purchases. By the end of the month, Sandbox offered ten lifetime pipeline runs, with local or permitted cloud execution and the applicable limits. Full-platform plans remain the route to more Agent Builder capacity, team seats and wider allowances.

Account management now shows the products you own, renewal or end dates, invoices and changes. A change review shows the new monthly total and any amount due today before you confirm payment. Model inference stays with your chosen provider.

These are the October launch terms, not a frozen price list. The current selector has since added Computer Control; check all current products, bundles and full-platform plans before subscribing.

Enterprise AI controls: clearer approvals and fewer surprises

Powerful tools are useful only when you can understand their reach.

October improved approval previews across chat, mobile and desktop. Social reply cards show the source item, not just the model's description of it. Odoo write previews explain the record and fields. Edited arguments are revalidated, canceled runs close their waiting approvals, and a decision made on one surface is reflected elsewhere.

Optional Koreshield scanning can inspect proposed actions for agents or conversations where it is selected. Shared run visibility and trace redaction were hardened. A public Data Processing Agreement also became available for business customers. These are concrete controls and documents, not a claim that every possible agent action is risk-free or independently certified.

There were everyday improvements too: mobile navigation and run views, more readable tool results, generated files that remain accessible from conversation history, better attachment handling, spreadsheet updates that refuse ambiguous keys, and outreach helpers that avoid repeating an already answered message. Those small boundaries matter when an experiment becomes a daily workflow.

Open-weight model workflows: choose one job to start

You do not need to adopt every Melaya product at once. Choose the task that currently leaves you copying an AI's answer into another app.

Frequently asked questions