What an AI agent can see on your KPI wall

Most integrations an AI assistant can reach are write-only. It can send something somewhere and, if nothing throws an error, assume it worked. That is fine for posting a message and surprisingly bad for anything you are going to look at later — because the assistant has no idea what it actually built, and neither do you until you open it.

We just changed that for Dashwall. An agent could already create a board, add tiles and push readings. It can now also ask what a board currently shows, which turns out to change what agents are useful for.

Reading the wall, not just writing to it

A Dashwall board is a set of tiles, each with a number, an optional target, and an optional comparison against an earlier period. All of that is what a person sees from across the room. It is now also what an agent sees: the latest reading of every tile, how far along it is against its target, how it moved, when it was last fed, and the history behind it.

> "How is the support board doing?"

Open tickets      31   ▲ 24% vs yesterday   ⚠ 3h ago
First response   4m 12s  ▼ 8% vs yesterday
Deploy pipeline  Operational

Three things follow from that, and only the first one is obvious.

• The agent can check its own work. Ask for a support board and it can confirm the tiles exist, are fed, and show what you meant — rather than reporting success and leaving you to discover the empty tile tomorrow.
• The wall becomes an input. “If the backlog is over fifty, draft a note to the team” is now a thing an assistant can do, because the backlog is something it can look up. The wall stops being only an output surface.
• You can ask about the board in words. Not everyone is standing in front of the screen. Asking an assistant what the wall says is a fair way to catch up.

A number that stopped moving should say so

There is a failure mode that arrives with any automated feed, and agents make it worse because they only run when something runs them. A tile fed by a script that quietly died on Friday still shows Friday's number on Monday — confidently, in very large type. That is worse than an empty tile, because nobody doubts it.

So a tile can now be told how often it expects to be fed. Past that interval, plus a little grace for a feed that is merely late, the wall stops pretending: the tile dims, it says how long ago it was updated, and a status tile falls back to “No data” instead of holding a green “Operational” from three days ago. The agent sees the same flag, so it can notice its own feed has stopped and go fix it.

Connecting without pasting a credential

Until now, connecting an assistant to Dashwall started the way these things usually do: open the dashboard, create an API key, paste it into a config file. That works, and it quietly excludes everyone who does not think of themselves as technical.

The hosted MCP server now speaks OAuth. The client registers itself, sends you to a Dashwall screen that names the application and lists what it is asking for, and gets its own token once you approve. Nothing is copied.

# The hosted server asks you to approve it in the browser
claude mcp add --transport http dashwall \
  https://mcp.dashwall.io/mcp

None of this is Claude-specific, and it would be a poor open protocol if it were. The same server is one config entry in Cursor and in Codex, which then run the approval themselves —codex mcp login dashwall in the latter case.

// Cursor — ~/.cursor/mcp.json
{ "mcpServers": { "dashwall": { "url": "https://mcp.dashwall.io/mcp" } } }

# Codex — ~/.codex/config.toml
[mcp_servers.dashwall]
url = "https://mcp.dashwall.io/mcp"

If you would rather hold a credential yourself, API keys are still there — and they got better at being handed out. A key can be read-only, so an assistant that only has to look at the wall never gets the ability to change it, and it can be given an expiry so the worst case ends by itself.

Why this matters more than another integration

Dashboards are usually built by whoever has an afternoon and the right access. That is why so many teams have a wall that is six months out of date: the wall is not the hard part, the loop of noticing, building, feeding and checking is. An agent that can do all four of those — including the checking — is a different proposition from one that can only create things.

The whole surface is on the agents page: the ten tools, both ways to connect, and what a credential is allowed to do. If you have not seen the basic version yet, start with letting an agent build the dashboard and then come back here.

Dashwall is launching soon. Join the waitlist and we will let you know the moment you can point an assistant at it.