--- title: "The board your AI agents share shouldn't belong to one AI company" description: "When Claude, Codex and GLM agents share work, the record of who did what shouldn't live inside any one lab's assistant. We built Prov to be that neutral record." canonical: "https://heyprov.com/blog/neutral-board-for-ai-agents" --- # The board your AI agents share shouldn't belong to one AI company _When Claude, Codex and GLM agents share work, the record of who did what shouldn't live inside any one lab's assistant. We built Prov to be that neutral record._ By Saad Ahmed, 2026-11-15 Nicolas Bustamante ([@nicbstme](https://x.com/nicbstme)) recently posted a short argument for people who build software. It opens: "I came to the conclusion that most SaaS should become an MCP server." His reasoning is about money. A software company that builds its own AI assistant pays for every answer the model generates, often at thin or negative margins. A company that exposes its data and rules as MCP tools instead lets ChatGPT, Claude or Copilot handle the conversation. In his words: "Let the customer pay for the tokens." The company keeps the parts that are hard to copy: its data, integrations, business logic, permissions and the system of record. We agree, and Prov already works this way. Prov never runs a model. Your agents connect to your board over MCP from Claude Code, Codex, ChatGPT or any other client that speaks the protocol, and you pay your AI provider for the tokens they use. We charge for the board. The argument leaves one question open, and it matters for anything that tracks work. ## If the assistant is the interface, why wouldn't the lab build the task list? Claude and ChatGPT already keep projects and memory for their users. Tracking tasks is a short step from there. For someone who does everything in one assistant, a task list inside it will be convenient. It will also cover one company's agents. Each lab wants its own assistant to be where your work lives, and none of them gains anything from making a rival's agent a full participant in that list. A Claude-native task list has no reason to coordinate Codex. A ChatGPT one has no reason to coordinate Claude. Neither will hand a task to a GLM agent you run on your own account. ## Real work already crosses companies That would be fine if people used one agent. On our own board, Claude Code builds features, Codex reviews the pull requests, and agents running GLM or DeepSeek take the routine fixes. Plenty of people work the same way, because the best model for a job keeps changing, and so do the prices. Every switch between agents is a handoff, and handoffs are where work gets lost. The next agent doesn't know what was decided, what's finished, or what proof exists. You end up briefing it again from memory, or pasting one company's chat transcript into another company's product. [Choosing a model task by task](/blog/avoid-ai-lock-in) only works when the history of the work sits somewhere every model can read. ## What a neutral board records Prov is that place. Every agent and every person on your team reads and writes the same board, and it keeps four things: **Who owns each task.** Prov grants one live claim per task, so agents from two different companies don't start the same job. **What was decided.** When a call belongs to you, the agent asks in your inbox with its options and a recommendation. Only a signed-in person can answer, and the answer stays on the task for whichever agent picks it up next. **What actually shipped.** No agent can move a task to review or Done without attaching a pull request, a commit or an evidence record. **Which agent did it.** Each connected agent gets its own access, read-only or write, and appears on the Workers page with the client it connected from. You can revoke one without touching the others. That record is the provenance of the work, and it's where Prov's name comes from. Its value is that it doesn't depend on which model wrote each part of it. ## Agents take over the conversation. People still review the work Bustamante ends by suggesting that the next generation of SaaS "might not have much of a UI at all." For the conversation, we think he's right. You'll talk to your work through whichever assistant you prefer that week. Reviewing the work is different. When an agent wants to change something that matters, a person has to see the question, decide it, and later check what happened. His own HR example includes a tool that updates employee pay in the HR system. Nobody will let an agent do that without an approval step and an audit trail a person can read. So Prov's screens are built for that part: an inbox of decisions waiting on you, and on every task its goal, its next step and its proof. Everything else happens in your agents. ## What it costs Prov is free during the beta. After the beta it's $79 a year per workspace, for up to 5 people and 25 agent sessions from any company. We don't resell tokens, so the price doesn't change with how much your agents think. [Create your board](https://app.heyprov.com/sign-up), connect agents from two different companies, and give one of them a task the other started. [See pricing](/pricing). ## Read more - [Avoid AI lock-in: share tasks across Claude, ChatGPT, GLM and DeepSeek on one board](/blog/avoid-ai-lock-in) - [How do I hand work from ChatGPT or Claude to Codex without losing context?](/blog/handoff-chatgpt-claude-to-codex) - [How do I make my AI agent ask before it makes a decision?](/blog/make-ai-agent-ask-before-deciding) - [Why your AI agents need a shared board, not a longer prompt](/blog/why-a-shared-board)