The web dashboard
The gateway has a web dashboard built in. There is nothing extra to install: start prism serve and open http://127.0.0.1:8130 in a browser.
If you set server.api_keys, the dashboard asks for a Gateway API key first. Paste one of your keys. Every request the dashboard makes uses the same key check as the API.
The pages
| Page | What it is for |
|---|---|
| Projects | Conversations, attached files and an optional agent, grouped per project |
| Playground | Chat with any model or auto, and see the routing decision |
| Models | Every model with its provider, context size, capabilities, speed and price |
| Routing | The default model, your rules in order and your Fallback chains, as the gateway loaded them. Read-only |
| Usage | Spent, Budget, Remaining, Requests, a daily cost chart and a table per model |
| Status | The gateway's version, uptime, mode and address, and the health of each provider |
The dashboard does not edit prism.yaml. Change the file, then restart the gateway.
Try a prompt in the Playground
Open Playground.
Choose a model, or leave it on
auto.Type a prompt and send it. The reply streams in, with the model that answered.
Select Explain route to see the decision without sending: the mode, the reason, the request profile and the Candidate chain.
Projects
Projects keep related chats and reference files together. Each project has its own standing instructions, attached files and, if you turn it on, an agent that can work with files.
Open Projects and select New project. Give it a Project name.
Start a chat. Chats are saved and stay in the project.
Use Attach files to add reference files the chats can use.
Open the project's Settings to add standing instructions for every chat in the project, then select Save settings.
Projects are stored as plain files in ~/.local/share/prism/workbench/. Chats go through the normal routing, so your rules, failover and budget apply.
The project agent
In a project's Settings, Agent tools turns on an agent that can list, read and write files, run shell commands and fetch web pages. It is Off — chat only until you switch it on.
Approval mode decides what waits for you:
| Mode | Behaviour |
|---|---|
| Always ask | Every tool call waits for you, including reads |
| Ask before changes | Reads run freely; writes and shell commands wait for you |
| Never ask | Tools run unattended |
Each tool call appears in the chat as it happens. A call that needs approval shows Approve and Deny.
DANGER
The agent's file tools stay inside the project's own workspace/ folder. Shell commands start in that folder, but they run as your user account and can reach anything your account can. Keep Ask before changes or Always ask, and read every command before you approve it. Do not use Never ask with a model or instructions you do not trust.
Choose a model that supports tool calls for an agent project, for example llama3.1:8b locally or a cloud model with the tools capability. Small coding models often describe a tool call in text instead of making it; the agent then behaves like a plain chat.