Workspace → Models (admins and owners) is where a workspace graduates from the managed Claude default to a model of its own: a small language model fine-tuned on the workspace's resolved conversations — including the tool calls behind them — served by Convoship or inside your own infrastructure. Guardrails, capability routing, budgets, and redaction are enforced by the Convoship runtime around any model, so governance never changes with the model choice.

Convoship vertical models — day one

A brand-new agent has no history to learn from. The catalog at the top of the page lists Convoship-published vertical models — telecommunications, IT support, finance, healthcare — that any workspace can deploy immediately: no training data required. Deploy one Convoship-hosted or as a self-host bundle, replay your own evals against it, and attach it. As your traffic accumulates, train a workspace-specific model and replace it.

Step 1 — build a training dataset

Real conversations — exports resolved conversations from a chosen agent (or all agents) and time window. Transcripts were PII-redacted when stored and are scrubbed again before training; an explicit data-use confirmation is required on every export.
Synthetic — for cold starts: each capability's example utterance expands into varied customer messages, answered by the live agent (tool calls included) across multi-turn exchanges — follow-up questions, slot answers, confirmations. Replies that didn't come from a real model are discarded automatically.
Tool calls are first-class in the training data: the model learns the exact wire format the runtime speaks, not just question-answer pairs.
Large corpora (tens of thousands of conversations) are generated as a managed batch job by your Convoship team.

Step 2 — train

Pick an open-weight base model (Qwen 2.5 7B or Llama 3.1 8B) and a model name, then start training. Convoship fine-tunes a workspace-specific adapter (LoRA); progress appears under Training jobs. The agent keeps serving traffic on its current model throughout.

Step 3 — evaluate before you trust it

Run evals on any live deployment replays your authored agent evals — the same cases that gate publishing — against the model's endpoint and stores the pass rate on the model card. Compare candidates on the cases that matter to you before switching any traffic.

Step 4 — deploy and attach

Host on Convoship — served from Convoship's model pool; goes live once provisioned.
Build self-host bundle — download a ready-to-run package (vLLM + your adapter) and run it inside your own network; conversation traffic then never leaves your infrastructure.
Use for this workspace — one click points every agent in the workspace at the live deployment, writing the same settings as Workspace Settings → AI Agent model. Verify with Test connection; switch back to the Convoship default at any time.
Optional safety net — Workspace Settings can enable an off-by-default fallback that retries failed custom-endpoint turns on Convoship's Claude. Leave it off if conversation content must never leave your endpoint.

Availability

The Models section appears once the fine-tuning platform is enabled for your environment — contact your account team. Without it, workspaces can still attach any OpenAI-compatible endpoint manually in Workspace Settings → AI Agent model.