# Free Models > Conserve budget by delegating suitable Sub-Agent work to free-of-charge Models (OpenRouter free variants, OpenCode Zen free tier) and by helping the user set up a free-capable Provider when none is configured. Use when the user asks to save costs or mentions free models, a low budget, or spending nothing, when many small independent subtasks can run on weaker Models, or before fanning out bulk Sub-Agent work. Do not use for critical, security-sensitive, or irreversible work that needs top-tier quality, and never as a license to downgrade the main conversation's own Model. - Skill: `vironnimo/free-models` (Agent Skill, multi-file: 3 files) - Install (CLI): `npx skillmds@latest add vironnimo/free-models` - Raw SKILL.md: https://api.skillmd.com/api/skills/vironnimo/free-models/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: AI & ML - Author: Vironnimo (https://skillmd.com/u/vironnimo) - Updated: 2026-09-17 - Page: https://skillmd.com/skills/vironnimo/free-models --- # Free Models Free capacity serves **delegated** work. You route suitable Sub-Agent Runs to free Models through the `subagent` Tool's per-Run `model` override (`/`, applies only to that Run). You never change the calling Agent's own Model, any stored configuration, or a Project default through this skill, and never do so silently. ## Non-negotiable rules 1. **Tools gate.** A Sub-Agent without tool calling is broken, not cheap. Propose only Models whose catalog record confirms tool support (`vbot model show `), and only ids that appear in `vbot model list` — a configured, usable Connection. 2. **No silent paid fallback.** When a free Model fails or returns unusable output, rotate to another free Model, wait out rate limits, or report the gap. Moving the work to a paid Model spends money and requires the user's explicit consent first. 3. **Quality boundary.** Ask before routing security-sensitive, architectural, irreversible, or otherwise high-stakes work to a weaker Model. Saving cents is never worth silently degrading an outcome the user cares about. 4. **Privacy boundary.** Free tiers are often paid in data — several free endpoints log inputs or train on them. Never put credentials, secrets, or tokens into delegated task text. Flag sensitive content to the user and prefer zero-retention, paid, or local alternatives for it. 5. **Rate limits are real.** OpenRouter free variants allow 20 requests per minute, 50 requests per day (raised to 1000 once the account has purchased at least $10 in credits, ever); expect HTTP 429 as routine. Spread siblings across different free Models instead of stacking parallel Runs on one, and treat exhausted quotas as a reason to defer work, not to fall back to paid. 6. **Verify before trusting.** The curated list in `references/free-models.md` is a dated hint. Free lineups churn weekly — confirm current free status and availability live before relying on any entry (recipe in that file). ## Workflow 1. **Inventory.** Run `vbot provider list` and `vbot model list --task chat` to see which Providers are usable. If no free-capable Provider is configured, stop here and walk the user through `references/provider-setup.md`; after setup, continue only if the user wants that. 2. **Pick candidates.** Read `references/free-models.md` for the curated snapshot and the live-verification recipe. Keep only entries that pass both checks right now. 3. **Match task to Model.** Delegate bulk summarization, triage and classification, structured extraction, formatting and boilerplate edits, first-pass review notes, and documentation drafting to free Models. Keep architecture decisions, security-relevant changes, final integration and verification of other output, and anything irreversible on the main Model. 4. **Delegate deliberately.** Write self-contained task text — goal, context, scope, constraints, expected result. Weaker Models compensate with precision, not intuition; spell out what a strong Model would infer. Respect the target Model's context window when the task carries large inputs. 5. **Handle failure.** On a failed or empty child Run, classify before retrying — rate limit (wait or rotate), capability miss (wrong Model for the job), or quality garbage (tighten the task or escalate). Never hammer a rate-limited Model in a retry loop. 6. **Report.** State which Runs ran free and on which Models, which stayed on the main Model and why, what failed or awaits consent, and the verified result. ## References Read `references/free-models.md` for the dated curated snapshot, current rate limits, and how to verify the live free lineup. Read `references/provider-setup.md` before walking the user through creating an OpenRouter or OpenCode Zen account and connecting it to vBot.