Scope
Pick the repo, goal, budget, and verification commands.
Bounded AI loops you can trust.
Forge runs a local plan, patch, gate, and review loop against your repositories. It uses your Claude, Codex, and DeepSeek lanes, stops at hard limits, and leaves a branch you can inspect before anything merges.
Demo Paused on cost cap $0.00048 spent
mission v3.2.0 release smoke
amaterasu / deepseek-pro
Not a SaaS. Not auto-deploy. Not a mystery agent.
Forge is a personal workshop for bounded AI development passes. It keeps the work local, routes each run through explicit checks, mirrors important state to Discord or phone push, and treats "ready" as inspection material rather than a shipped change.
How the loop works
Every run has a budget, a branch, proof commands, and a stop condition.
Run
Forge plans, patches, gates, and reviews on an iteration branch.
Halt
Cost caps, rate limits, failed gates, and human questions stop the loop.
Inspect
You get the branch, logs, verdict, costs, and next action.
What v3.2.0 brings into view
Mission control, cost awareness, and autonomous recovery without giving up human review.
Mission autopilot
Chains scoped mission actions, then stops cleanly when a gate, edge case, or human decision needs attention.
DeepSeek lane
Routes cheaper work through DeepSeek with tracked spend, live cost updates, and low-cap halt behavior.
Persona-aware voice
Local Amaterasu, Tsukuyomi, and Raijin records shape prompts without depending on Kanade being online.
Always-on runner
NSSM keeps the dashboard and orchestrator available after terminal windows close.
Auto-resume
Rate-limit pauses can schedule a Windows Task Scheduler resume instead of losing the run.
Phone push
ntfy.sh can mirror run state changes to a phone without adding a paid notification service.
Local persona voices
Forge can shape runs with local persona records, even when Kanade is offline.
Clear the path, preserve the proof, and bring the branch home warm.
Slow the room down. Name the risk before the next command moves.
Hit the build hard, report the thunder, leave no hidden failure mode.
The dashboard is an inspection bench.
This preview shows the shape of the workbench: active missions, resumable runs, visible costs, persona controls, and run history that keeps the proof next to the patch.
Forge runs / settings / inbox
active mission forge-mission-real-smoke
halted · verdict pending · amaterasu / codex
forge-v320-mission-smoke 2/2 iterations
ready $0.00 forge-v320-deepseek-smoke 1/1 iteration
ready $0.00048 forge-v320-cost-cap-smoke cost cap
awaiting-human $0.00001 cap Personal tool, public demo.
Forge is local-first, self-hosted, and built for one operator's real workflow. This page is a public-facing preview of the idea, not a product funnel, account system, installer, or telemetry pitch.