Chat with an Agent, publish to Slack, or embed a website chat

Hosted work requires a configured model provider and workspace funding; purchases require configured payments.

An Agent remains useful with your existing agent through Fermium MCP. Optional hosted work lets Fermium run the model for a task, using the Agent's workspace credits. You can use it in Chat, ask for the same work through MCP, or configure Slack and website deployments from Publish.

Publish an Agent to Slack

Open Agent → Publish → Slack to let your team use the Agent in Slack. Slack deployments need a configured model and workspace AI credits. The Linked Fermium accounts audience also requires hosted chat to be enabled for the Agent.

1. Under Bot identity, choose the shared Fermium app for quick setup, or Dedicated Slack app for the Agent's own name and @mention. The dedicated option provides a manifest and credential setup steps.
2. Install the app in your Slack workspace and choose the channels it may use.
3. Review the audience. Linked Fermium accounts uses each linked person's access. Broader channel access requires explicitly reviewed and published knowledge and tools; inviting a bot does not grant access to private documents.
4. Test the connection, review the deployment and its limits, then choose Publish Slack Agent.

With the shared bot, people select the Agent they want to use. A dedicated bot is bound to its Agent. Tool policies and required approvals still apply. You can pause the deployment to stop new use; removing the Slack installation is a separate step.

See Agents for knowledge, source permissions, and tools.

Start a conversation

1. A workspace owner or admin enables hosted usage under Settings → AI credits and funds the workspace. Available balance, purchase limits and model rates are shown there; rates depend on the configured catalog.
2. An Agent owner or workspace admin enables its hosted runtime and chooses an available model in Chat settings. Choose team work (private unless explicitly shared), personal work, or isolated conversations (sharing disabled).
3. Open Chat and start a conversation. It is private by default, including in an Agent other workspace members can access.
4. Ask a question, inspect the cited sources and saved tool steps, and follow up in the same conversation. Keep the default Research scope for analysis.

Try: “Compare the client brief with our delivery notes. Identify three gaps, cite the evidence, and keep a draft summary in this conversation.”

Research can read authorized Agent context and sources. Connected tools must be explicitly designated for research by a connection manager. Merely naming a tool “search” does not make it permitted. The regular policy and approval rules still apply.

Request actions explicitly

For a task that changes an external system, select Actions and describe the intended changes. This scope allows the model to request actions; it does not waive tool rules or required approval.

Try: “Compare the client brief with the connected project, explain what's missing, and draft issues. Create the issues once the required approval is granted.”

When a run waits for approval, an authorized reviewer decides the exact action in Audit. The original requester sees Waiting for a reviewer until a decision, then an explicit continuation action. Approval does not execute the action by itself. The waiting period does not keep a model generating. A shared-session collaborator cannot inherit the original requester's pending approval.

Runs show progress, credit use and an outcome. Cancellation stops future steps; it cannot undo an action already sent to another system. If an action's outcome is unknown, inspect that system before requesting the action again.

Resume or share a conversation

Use the searchable conversation list to find work marked In progress, Needs follow-up, Done or Archived. A completed model response leaves work In progress until you or an authorized agent changes its work status. Its title, summary and next step help you find where to continue. You can resume the same conversation from the app or MCP. Restore an archived conversation before continuing it; archive is not delete.

Open Share in the conversation header. The session owner can deliberately grant View or Continue access to a person, group or the workspace, then select Save conversation access. Added and removed entries remain unsaved until that action. Sharing covers the full conversation and its working documents. Recipients also need access to the Agent and the source documents used in the conversation. Continue runs new work as the current person; it does not transfer credentials, management rights or another person's pending approval.

Workspace audience enables deliberate sharing. Personal/customer audience is for private work; revoke existing session grants before changing to either audience. Wait for or cancel active work before changing a session's sharing or status. Revoking a grant prevents subsequent access; it cannot retract content someone already downloaded.

Agent notes and ordinary memory files still belong to the Agent's audience. Keep private conversation details in the session rather than writing them into the shared journal.

Propose useful findings for review

Try: “Draft a reusable memory document from this conversation. Propose it for the Delivery Agent, and wait for review before sharing it.”

The agent creates a draft containing the proposed document and its destination. It remains within the session until its owner submits it for review. Submission shares that proposed document with eligible Agent reviewers; it does not share the whole private conversation.

A person with the required write access chooses Approve and publish or Reject. Acceptance publishes only the proposed document to the chosen Agent/workspace destination; workspace documents go into General and are available to all workspace members, which can be broader than the Agent access list. An authorized reviewer may also be the author; this is distinct from independent tool approval. Agent memory also gains an index reference. Agents can draft a proposal, but cannot submit or approve it through MCP. Existing direct writes to ordinary Agent memory keep their established permissions; this review flow handles promotion out of private sessions.

Put a knowledge chat on a website

The first deployment surface is public knowledge chat. It answers using a separate body of knowledge you explicitly approve for public visitors. It has no access to private Agent context, private sessions or connected upstream tools.

1. Open Publish as the Agent owner or a workspace admin.
2. Give the chat a name and greeting, choose a model, and enter the knowledge suitable for public visitors. Private Agent files are not automatically included.
3. Choose exact embed origins, a lifetime deployment budget and a per-message ceiling. Save private draft leaves any live chat unchanged.
4. Use Private preview to test a visitor question. The test conversation is accessible only to you, uses workspace credits, and cannot use private Agent context or tools. It does not consume the public deployment's lifetime budget.
5. Choose Review and publish, compare current and proposed knowledge, confirm the audience and spending, then publish. Copy the iframe into your site.

The card shows publication version, time, publisher and a content fingerprint. Draft changes remain unpublished until reviewed. Pause public chat stops new requests and further hosted steps. It cannot retract answers already delivered. These are separate approved snapshots; edits to Agent files never silently change a public chat.

An authorized agent can prepare this setup with prepare_brain_deployment and give you its review link. The result is always disabled until a person reviews the prepared knowledge and explicitly publishes it. Ask: “Prepare a chat from our approved FAQs and send me the deployment review link.”

Use an exact origin such as https://www.example.com, without a path or trailing slash. The origins control where the browser permits embedding; they do not make published knowledge private or authenticate a customer. Anyone with the enabled chat URL can open its public experience.

Visitor conversations are separate. The website snippet contains no privileged Fermium or provider credential. Pause the deployment to stop new use. Authenticated customer portals with private account data, deployed upstream actions and voice interfaces are outside this initial implementation.

Ask from your existing agent

Your agent can create or select an authorized session, call run_brain, then poll get_brain_run for progress. It reads the answer and working documents from the session. Individual Agent, file, search and connected-tool calls remain available.

Try: “Ask Fermium's Delivery Agent for a cited gap analysis. Keep the work in a private session and tell me what remains unresolved.”

See Agent sessions and hosted work for exact parameters, approval continuation, session sharing and contextual search. See Billing for workspace funding.

Current limits

Hosted runs are bounded and may stop for credit limits or a step limit. Saved work remains available; use a follow-up to continue after inspecting the outcome. Provider failure does not imply completed actions or free usage. Search uses keyword and literal matches with expandable surrounding context; semantic/vector search is not part of this implementation. Available model choices depend on operator configuration, and token/storage limits still apply to file-based context.

Your account export includes your accessible owned conversations and their saved entries. Agent deletion removes its conversations, while account purge cleans up owned runtime data. Archive only moves work out of the active list; it does not erase it. See export and deletion.

What the model receives

Hosted work sends the request, relevant conversation context, authorized Agent/source content and tool results to the configured model provider to produce the answer. “Private session” controls who can open it in Fermium; it does not mean inference happens without a model provider. Connected actions also send their approved arguments to the upstream service. Required policy assessment can use a separately configured provider and may block an action if that assessment is unavailable.

Hosted private search queries and detailed results remain in the session; shared activity records the steps without those query details. Individual Agent tools used directly by an external agent retain their normal journal behavior.

Public chat uses only the knowledge you published for that deployment and its visitor's own conversation. Provider retention and contractual settings require separate configuration; this local-review guide does not promise a particular provider retention period or training arrangement.