Models & inference

Objective is model-agnostic. Objectiv stores endpoint configuration in app settings—keys are not checked into your repository.

Local Ollama

Install Ollama on the same machine as Objectiv (or reachable on LAN).
Pull a code-capable model (team choice—balance quality vs VRAM).
In Objectiv: Settings → Objective → set base URL to your Ollama host.
Run a small agent task to validate latency and context size.

Best for: air-gapped networks, laptops, on-prem GPU boxes.

Bring your own key (OpenAI-compatible)

Point at any gateway that speaks chat completions compatible with OpenAI’s API shape:

Commercial providers (OpenAI, Anthropic via proxy, etc.)
Corporate inference gateways
Cloud GPU endpoints you operate

Rotate keys in Settings without rebuilding the desktop app. Build-time env vars exist for OEM deployments—see objectiv/app/.env.example.

Enterprise policy (roadmap)

Seat licenses already gate comms tools. Future JWT claims may include allowed model endpoints or org-mandated gateways so admins can block arbitrary URLs on managed machines.

Quality & fine-tuning

Repository doc OBJECTIV_MODELS.md covers which models the team evaluates, QLoRA SFT/RL plans, and offline G1 milestones. Your day-to-day choice still lives in Settings unless org policy overrides.

Privacy checklist

Local Ollama: prompts and code snippets sent only to localhost.
Cloud API: treat prompts as sensitive—use corporate gateways with logging policies you accept.
Approve gate: reduces risk of unreviewed writes regardless of model host.