Finta assembles a quality-scored catalog of 2,500+ agents, skills and commands into a ready-to-run config for 19 AI assistants — then forges self-verifying agent loops with guardrails, on-disk memory and human escalation.
A wizard for the setup, a forge for the execution layer. Both run as static pages — open a file, get to work.
Every agent, skill and command carries a 0–100 score from source reputation, content depth and curation tier. Only items ≥ 65 surface by default.
Claude Code, Cursor, Copilot, Windsurf, Codex, Gemini, Aider and 12 more — the same 5-minute wizard generates the right config format for each.
New: forge agents that run a verified loop — observe, decide, execute, verify — and stop on a provable goal instead of drifting forever.
Iteration caps, stagnation detection after 3 flat rounds, time/token budgets and rollback on regression — enforced in code, not prose.
Permanent memory, an incident buffer purged on resolution, and an exchange bus that sub-agents read and write to coordinate.
Five escalation triggers — ambiguity, irreversible actions, sensitive data, subjective judgment, exhausted budget — hand control back to you.
A persona file tells an agent who to be. A loop tells it how to finish. Loop Forge generates the full execution architecture — orchestrator, config, memory, scoring grid and a runnable Python loop runner.
Doubt above 10%? The agent asks instead of guessing. A bug? One diagnosis, one atomic fix, max 3 retries, then a regression test before it counts as resolved.
Turn-by-turn, goal-driven, temporal, proactive, or simplified with human validation — picked by a 3-question decision tree, not a single hardcoded mode.
Deterministic criteria (tests pass, CI green) and soft criteria (matches the spec, validated once with you) — never a blind pass/fail boolean.
Single agent, parallel fan-out, sequential chain, or mixed — with an exchange bus file that every sub-agent reads and writes.
Not just markdown: a ZIP with AGENT.md, loop.config.json, scoring grid, memory files, structure validators and a Python loop runner with the guardrails built in.
Pick a profile — Fast, Advanced or Deep — and describe your stack, team and workflow.
Finta selects the best-scored agents, skills and commands for your context and builds a bootstrap prompt plus a file-tree preview.
Paste the prompt into your AI assistant once. It installs the config without overwriting anything already in your project.
Twelve-plus sources, each scored and tiered — plus Finta’s own sport-science and coaching agents you won’t find anywhere else.
Open the wizard, answer a few questions, paste one prompt. Your assistant does the rest.