Forge

Forge is the console’s chat. Open Forge, the first item under Build, and ask it about your workspace: which agents and workflows you have, how one is set up, what a run did, why an answer missed. Forge looks things up rather than guessing, and it acts as you: it sees only what you can see.

Talk to Forge, or to an agent

The Talk to menu under the message box chooses who answers your next message: Forge, or any agent or workflow you can use.

When you pick an agent or a workflow, it answers in a conversation of its own, exactly as its users get it, with its own instructions, tools and knowledge. It never sees what you said to Forge. Its replies still appear in your Forge conversation, under its name, with a line marking where the conversation changed hands. Pick Forge again and ask about the exchange: Forge can read it.

That conversation also appears in the agent’s own Playground history.

Model and attachments

When you talk to Forge, two more menus sit beside Talk to:

• Model chooses the model Forge runs on: the platform default, or any chat model your workspace lets you use.
• Attach gives Forge more to work with: MCP servers your operator allows for untrusted use, and knowledge collections you can read, which Forge can then search.

Both are checked again every time Forge runs, and they stay with the conversation. With an agent picked, the agent runs on its own model and attachments.

Trying an agent

Ask Forge to try an agent or a workflow, for example “try billing with ‘how long do refunds take?’”, and it asks you first. Once you allow it, the agent runs in a new conversation of its own. Forge’s reply then shows a card with the outcome: whether it finished, what it said, which tools it called, the cost of a workflow run, and links to the conversation and its traces.

A run spends model calls the same way a Playground run does.

Changing your workspace

Forge can create an agent, change an agent’s settings and instructions or its AGENTS.md, write your workspace’s skills, and create or change a workflow. It can also create a plugin, put new agents, workflows and skills into it, and import a plugin whole. It always asks first: you see the change against what is saved now, and you allow or deny each one. There is no “always allow” for a change.

How a change is shown depends on what it is. An agent or a skill is a diff. A workflow is a list of what is new, changed and removed, a canvas of the proposed workflow with its new and changed boxes marked (removed boxes are in the list, not on the canvas), and its YAML as a diff. A plugin’s description and version are shown as they are now and as proposed. A plugin import shows what it does to each agent, workflow and skill, then each file against the plugin as it is now.

If someone edits the same agent, skill, workflow or plugin while you are reviewing, Forge does not overwrite their edit: it reads the new version and proposes again. An import is held to the same rule. If anything it covers changed since Forge planned it, or if it would remove or change members you cannot see, it is not imported, and the review warns you before you allow it. What Forge creates is untrusted, and it never imports over a trusted plugin.

Deny with a reason, or just type what you want instead, and Forge revises its proposal. A change you allow shows as a card in Forge’s reply that links to its editor, a workflow to its canvas and a plugin to its page, and it shows in the audit log as yours, made through Forge.

Putting work to use

Forge can also put an agent or a workflow to use. It can:

• create, change and revoke a website embed;
• create, change, pause, resume and delete a schedule, and pause, resume, rename or delete a webhook (Triggers and automation);
• run an evaluation of an agent or a workflow; and
• generate a golden set for a collection.

Each is approved every time, a revoke and a delete included, and there is no “always allow” for any of them. An evaluation and a golden-set generation spend model money, and a generation always has a spending cap that you set.

The review shows each against what is saved now:

• An embed is a table of its name, allowed sites, greeting, starter prompts, limits and theme, as they are and as proposed. A revoke says that every site using the embed stops working and that it cannot be undone, and an embed that is revoked already says that revoking it again changes nothing. A revoked embed cannot be changed, and the review says so.
• A schedule is a table too, with the schedule in words and its next runs. A schedule runs as the person who last created, changed or resumed it, so the review says when saving or resuming makes you its owner. A delete says that its run history goes with it.
• An evaluation names what it runs against and says that it costs model spend.
• A generation shows a fresh cost estimate and the two models next to your cap, and warns when the estimate is above the cap, in which case it would stop at the cap with fewer items.

As with any change, the review warns you when an embed or a schedule has changed since Forge read it: if you allow it then, Forge saves nothing, reads the new version and proposes again.

After you allow one, Forge’s reply shows a card. An embed’s links to its page, and a new embed’s also carries the snippet to paste into your site; a schedule’s links to the agent’s or workflow’s Triggers tab. An evaluation’s says whether it passed, with its scores, and a generation’s says how far it got and what it spent. Forge waits up to five minutes for either; if one is still running by then, the card says so and where to look. The items a generation writes wait for you to review them on the collection’s Golden set tab.

Webhooks and API keys

A webhook’s address and an API key’s value are secrets, shown once, so Forge never creates either: they would pass through the conversation. It gives you a link to the console’s own form, filled in. For a webhook the name and the message are filled in; for an API key, the label and the role. You choose the webhook’s signing secret yourself, nothing is created until you click Create, and you copy the address or the key from the console, not from Forge. A key cannot hold more than you do: if Forge asks for a role above what you may grant, the form opens with the highest role you may.

What Forge does not do yet

Forge cannot create a webhook or an API key, set or change a webhook’s signing secret, rotate a webhook’s address, or regenerate an embed’s key: each would put a secret in the conversation. For those, use the console, which Forge links to. It also cannot remove the notification list from a schedule; do that on the Triggers tab.

History

History lists your Forge conversations, and New conversation starts another. Each one is yours alone.