--- title: "Cerevisor 3.8: a proper chat for local AI agents, and help that sees your screen" slug: cerevisor-v38-local-ai-agents-lasting-conversation date: 2026-10-08 excerpt: "In Cerevisor 3.8.0, talking to an agent is a proper chat, and the built-in help, Need a hand?, now knows what is on your screen. How both work." featured_image: "https://bbtxujdxvidaghmhxkqs.supabase.co/storage/v1/object/public/generated-images/blog-1791445485903-cerevisor-v38-local-ai-agents-lasting-conversation.webp" featured_image_alt: Illustration of a tidy wooden desk with an open paper notebook held by a ribbon bookmark, a small brass padlock resting on its page, and a row of closed blue notebooks standing beside it. author: Kinan Hamwi author_url: https://www.linkedin.com/in/kinanhamwi/ canonical_url: https://cerevisor.com/blog/cerevisor-v38-local-ai-agents-lasting-conversation updated_at: 2026-10-08T07:58:05.555065+00:00 --- # Cerevisor 3.8: a proper chat for local AI agents, and help that sees your screen By the founder of Cerevisor. Checked on 8 October 2026 against the source code of Cerevisor 3.8.0. TLDR In Cerevisor 3.8.0, released on 8 October, talking to an agent is now a proper chat: replies stream in as they are written, the box is never locked, and the conversation is still there after a restart. The built-in help, Need a hand?, now knows what is on the screen when a question is asked. Every chat message is held to the same checks as a run, the first launch connects the AI tools already installed, and the run report opens by itself. Below: how it works, why, and where it still stops. Talking to an AI agent in Cerevisor now feels like the chat apps everyone already knows. The reply appears as it is written, the box never locks, and the conversation is still there on Monday. And when something on screen makes no sense, the built-in help, called Need a hand?, already knows which screen that is. Those are the two headlines of Cerevisor 3.8.0. Cerevisor is a desktop app for people who do demanding work on a computer with AI. AI agents are laid out on a canvas as a workflow, and each one hands its files to the next. An agent here is an AI model given a task, some tools, and room to take several steps by itself. A workflow is a set of agents wired in the order they work, and a run is one pass through it. The app, the files and the record of every run stay on the computer, whether the model is paid for online or runs on the machine itself. That is what I mean by local AI agents: they are run from this computer, under rules set on this computer. Every change is listed in the [3.8.0 release notes](https://cerevisor.com/releases/v3-8-0). This post is the reasoning behind them. --- ## Talking to an agent is now a proper chat Every agent on the canvas has its own chat, and 3.8 rebuilt it around how people really type. **Replies arrive as they are written.** The text streams in, formatted, with a small card for each tool call. A tool call is one use of a tool by the agent, such as reading a file or running a command. A closed card is one line: how the call ended, which tool, and how long it took. Open it for the details. **The box is never locked.** It grows while typing, up to 6 lines under the agent’s card and 10 in the chat window. Any file can be picked, pasted or dropped onto it. Stop is there for the whole of a reply. **Two tabs, one box.** The chat opens on Activity, the live log of what the agent does in a run, with a Chat tab beside it and the same box under both. Sending a message shows Chat. A new run brings Activity back. **One status line says what a message would meet right now.** The chat status line in Cerevisor 3.8.0 It saysWhat the agent is doingWhat a message does ReplyingWriting a reply in this chatJoins the reply, or waits in a queue SteeringWorking in a run, and able to take a messageSteers the work in progress Working in the runWorking in a run, and not able to take a messageStays in the box until the agent is done Waiting for its turnThe run has not reached this agent yetIs added to what the agent gets when its turn comes Which row applies depends on what the agent runs on. A model Cerevisor calls itself takes a message in the middle of a reply. So does Claude Code, Anthropic’s coding agent. Codex (OpenAI’s coding agent), Cursor, Grok Build and Antigravity are separate apps that cannot, so the message waits in a queue. A queued message can be edited, taken back, or sent with Send now. --- ## One conversation per agent, kept on this computer Each agent has one conversation, and it lasts. It is still there after a restart, after a Stop and after later runs. It is written to a folder named agent-conversations inside Cerevisor’s own folder in the home folder, and never into the [workflow file](/blog/cerevisor-v34-v35-share-and-steer), so sharing a workflow does not share what was said to its agents. Anything that looks like a password or a key is blanked out before a line is saved. Reset starts a clean conversation without deleting the old one, which moves to Earlier conversations. A message to an agent that has finished now continues the conversation. Before 3.8 it ran the agent again. And the chat follows the run: every agent that finishes in a run gets a line for that run in its conversation, with how it ended and the start of its result, cut at 600 characters. So the first question after a run lands on an agent that knows what it did. --- ## Why a chat message runs under the rules of a run If a run has to ask before it writes a file and the chat does not, the real work quietly moves into the chat. So in 3.8 a message is not a side channel. Each one starts a small run of that single agent, through the same entry point every run uses. The agent answers with the tools, permissions and instructions on its card. The approval setting, which is the choice of how often Cerevisor asks before an agent acts, applies as it does in a run. Tool hooks still fire. A hook is a script of the operator’s own that Cerevisor runs when an agent uses a tool. Each reply is recorded, and listed in the System Log as a Chat turn. One difference is deliberate. Hooks written for the start and the end of a workflow run do not fire for a chat message, or a hook meant to announce a finished run would go off after every question. > A chat box that skips the checks is the fastest way around every control in the product. In 3.8 a message is a small run, so it cannot skip them. A lasting conversation also changed one safety rule. Once an agent has read content from outside, a web page for example, Cerevisor is more careful with what that agent sends out. In a chat, that care now lasts until Reset, because the agent’s earlier replies stay part of what it reads. So after a research conversation, I reset before asking the same agent to send anything anywhere. --- ## Need a hand? now knows what is on the screen Need a hand? is the help built into Cerevisor. In 3.8, each question arrives with a short description of the screen: the open world (a world is one canvas that holds several workflows), the workflow in focus, the selected agent, and that workflow’s last run and error. “Why did this fail?” is now a complete question. The help pages come back one section at a time, so no help text is out of reach, and it may look things up for as many as 10 rounds before it answers. It can explain a run, list a workflow’s problems, describe an agent, and explain how data moves between agents. What it is sent is narrow on purpose. The description holds names and states only, never memory and never an agent’s instructions. It rides on the newest question alone and is not saved with the help conversation. An error line is cut at 300 characters. Square brackets and double quotes are taken out of every name before it is sent, so a workflow with a crafty name cannot pass itself off as part of the app’s own description. It still cannot build, edit, save or run anything. Asked for a change, it explains how to make it. The screen details go to the default model provider with each question, the provider being the service an agent’s model comes from, and the [help assistant page](https://cerevisor.com/docs/guides/help-assistant) says exactly what is sent. --- ## First launch and connected tools: less setup, the same checks **First launch connects what is already installed.** On a computer that has never had a provider in Cerevisor, the first launch looks for Claude Code, Codex, Grok Build and Ollama, a program that runs open models on the machine itself. Each one that is installed and signed in is connected without a click, and one notice names it. It happens once per computer and without a test, so an expired sign-in shows on the first run. The steps are in the guide to [connecting a model provider](https://cerevisor.com/docs/getting-started/connecting-a-provider). **Claude Code and Codex agents can use connected tools.** A connected tool is an outside service an agent can call, such as GitHub or a calendar. It is attached through an [MCP server](/blog/harness-how-to-setup-mcp-server-agent-workflow), a small program that speaks the common protocol AI tools use to call other software. Before 3.8 these agents were shown connected tools they were never handed, because the two apps make their own tool calls. Now each such call is made by Cerevisor, through the same ten stages any other tool call goes through, the approval and the hooks included. Cerevisor cannot see what such an agent reads with its own tools, so a call to a remote server that is not marked Trust this server asks first when someone is watching, and is refused in a scheduled run. Other tools are tightening the same handover. From the Claude Code changelog for version 2.1.292, dated 6 October: > "Fixed an MCP tool with a name longer than 128 characters making every request fail; that tool is now left out and an MCP error names it" Claude Code changelog, version 2.1.292, 6 October 2026 **A sign-in that could not work.** In release week, signing in to GitHub’s server with the browser stopped on this error: “Incompatible auth server: does not support dynamic client registration”. The fix shipped on 7 October: GitHub’s card now asks for a token and says where to make one. Setup is covered in the [MCP servers guide](https://cerevisor.com/docs/guides/advanced/mcp-servers). --- ## The run report opens by itself, and shared workflows join the open world When a run started in the window finishes, its report opens in Cerevisor. Runs started by a schedule, a script or a phone still write the report. They do not open it. The share menu has Import shared workflow, which adds a shared file’s workflows to the open world as new ones. None of the sender’s approvals travel with it. --- ## Known limits in Cerevisor 3.8.0 These come straight from the release notes. - **Chat during a run:** an agent working in a run on Codex, Cursor, Grok Build or Antigravity cannot be given a message until it is done. - **The help assistant** sends more text with each question, so a question costs a little more, and a small local model may not have room for it. - **Connected tools:** Cursor, Antigravity and Grok Build agents cannot use them yet. They were tried on Codex 0.160 only, and Codex 0.161.0 came out on 7 October, the day before this release. - **Tools set up inside Claude Code or Codex themselves** still load, and those calls are not Cerevisor’s to check. - **Browser sign-in for connected tools** works only with services that let an app register itself. GitHub and Slack do not. - **Claude Code on a Mac** is not found at first launch, because its sign-in is kept in the Keychain. - **Chats from before this version** are not carried over, and the run report is always in English. - **Windows installers are not yet code-signed and macOS builds are not notarized.** --- ## What I would do first after updating I would pick one agent whose work I know well and talk to it for a day. Ask why it made a choice, type a second message while it is still answering, close the app, come back, and ask a follow-up. Then filter the System Log in Analytics to Chat turn. If that list reads like a fair record of the day, the chat is doing its job. Next I would open a workflow with a run that went wrong, press Need a hand?, and ask why it failed without describing anything. The answer should name the agent and the error. Existing installations get 3.8.0 through the built-in updater. For a first install, the installers for Windows, macOS and Linux are on the [download page](https://cerevisor.com/download). #### Sources - [Cerevisor 3.8.0 release notes](https://cerevisor.com/releases/v3-8-0) - Cerevisor, 2026-10-08 - [Claude Code changelog, version 2.1.292](https://github.com/anthropics/claude-code/blob/main/CHANGELOG.md) - Anthropic, 2026-10-06 - [Codex release 0.161.0](https://github.com/openai/codex/releases/tag/rust-v0.161.0) - OpenAI, 2026-10-07