Open-source agent orchestration framework. V11's proven patterns, any LLM, pure Python. v0.1.0 Amazon Nova Hackathon
# Plan a new project
forge plan "weather dashboard" --model nova-lite
# Build it (agents execute in parallel waves)
forge build --model gemini-flash
# Deploy to a live URL
forge deploy --domain weather.herakles.dev
# Check status
forge statusNova Forge replaces Claude Code's closed-source agent runtime with a ~300-line Python tool-use loop that works with any LLM supporting function calling.
Key insight: V11 built 89 services on the Hercules platform. Nova Forge proves those patterns are model-portable — we run them with Amazon Nova and build live apps.
| Alias | Full Model ID |
|---|---|
claude-haiku | anthropic/claude-haiku-4-5-20251001 |
claude-sonnet | anthropic/claude-sonnet-4-6-20250514 |
gemini-flash | openrouter/google/gemini-2.0-flash-001 |
gemini-pro | openrouter/google/gemini-2.5-pro-preview |
nova-lite | bedrock/us.amazon.nova-2-lite-v1:0 |
nova-premier | bedrock/us.amazon.nova-premier-v1:0 |
nova-pro | bedrock/us.amazon.nova-pro-v1:0 |
| Formation | Roles | Use Case |
|---|---|---|
bug-investigation |
3 | Root cause is unknown. Three parallel investigators with distinct strategies (Ba... |
code-review |
3 | PR review, code quality assessment, pre-merge checks. Three reviewers in paralle... |
feature-impl |
4 | Adding features to an existing project. Most common formation. Backend and front... |
lightweight-feature |
2 | Small, single-layer features (frontend-only OR backend-only). 4-8 tasks. <4 hour... |
new-project |
3 | Greenfield project setup, scaffolding + initial implementation. Architect output... |
perf-optimization |
2 | Performance work — profiling, optimization, regression testing. Sequential subag... |
security-review |
3 | Security audit, threat modeling, vulnerability assessment. Threat modeler and sc... |
single-file |
1 | Small, focused file edits. 1-3 tasks, single file. <2 hours total. Examples: fix... |
Phase 1: Planning — ForgeAgent + smart model generates spec.md
Phase 2: Decomposition — ForgeAgent breaks spec into tasks.json
Phase 3: Execution — Wave-dispatched parallel agents build the project
Phase 4: Gate Review — LLM reviewer produces PASS/FAIL/CONDITIONAL
Phase 5: Deploy — Docker build, nginx config, health check