Tale connects to AI models through providers — any OpenAI-compatible endpoint qualifies, including local runtimes like Ollama, vLLM, and LocalAI. Running agents against a local provider keeps prompts, completions, and knowledge-base context inside your own network; nothing reaches a hosted model vendor. This tutorial walks through adding Ollama as a provider, attaching a model to an agent, and switching it on.
You need Admin access. A working Ollama or vLLM server reachable from your Tale instance is the only external prerequisite — see their respective install guides.
Step 1 — Start the local runtime
For Ollama on the same host as Tale:
ollama pull llama3.3
ollama serve
http://localhost:11434 by default. Confirm it responds:
curl -s http://localhost:11434/api/tags | jq '.models[].name'
localhost — on Linux, use http://host.docker.internal:11434 with an explicit Docker extra-host, or the host’s LAN IP. See Self-hosted configuration for the networking options.
Step 2 — Add the provider in Tale
Navigate to Settings > Providers and click Add provider. Fill in:
| Field | Value |
|---|
| Name | ollama-local (slug, used internally) |
| Display name | Ollama (local) |
| Base URL | http://host.docker.internal:11434/v1 |
| API key | Any non-empty value — Ollama does not auth |
| Field | Value |
|---|
| ID | llama3.3 (must match Ollama exactly) |
| Display name | LLaMA 3.3 |
| Tags | chat |
http://vllm.internal:8000/v1), and the model ID must match the --served-model-name flag you started vLLM with.
Step 3 — Add the model to an agent
Open Agents, pick the agent you want to run on the local model, and open its JSON at TALE_CONFIG_DIR/agents/<slug>.json. Add the model ID to supportedModels:
{
"supportedModels": ["llama3.3"]
}
{
"supportedModels": ["ollama-local:llama3.3"]
}
Step 4 — Test from chat
Open Chat, pick the agent, and ask anything. In Settings > Usage analytics or the agent’s conversation history, confirm the request was served by ollama-local — the model ID appears in the thread metadata. Latency will be higher than a hosted frontier model; that is expected on most hardware.
If Tale falls back to a different provider, either the model ID does not match Ollama’s list, or supportedModels still contains a frontier-model entry that takes precedence — remove or reorder.
Step 5 — Wire it into the must-have tutorials
Both admin must-haves benefit from a local provider:
- Office Agents — the add-in hits Tale; Tale routes to the local model. No change to the add-in side. See Word & Excel add-in.
- Meeting transcription — Meetily already runs Whisper locally; adding a local provider closes the loop so the summarisation LLM is also local. See Meeting transcription.
Troubleshooting
- Tale cannot reach Ollama from Docker —
localhost inside the Tale container is not the host. Use host.docker.internal (Docker Desktop), the host LAN IP, or put Ollama and Tale on the same Docker network.
- 404 on model — model ID is case-sensitive and must match the name
ollama list prints.
- Empty or very short answers — default Ollama context window is small. Pull a larger-context variant or override
num_ctx in the model’s Modelfile.
- Encrypted API key required — if you edit provider files directly, the API-key file must be SOPS-encrypted. Setting the key via the UI handles encryption for you; see Providers — SOPS-encrypted secrets.
Last modified on April 21, 2026
