FAQ
General
What is Aether Forge?
A spec-first agent builder framework. You write a strategy in plain English, run one CLI command, and get a typed, governed, production-capable agent with wallet, memory, MCP tools, and A2A communication.
Who is this for?
- DeFi engineers building autonomous trading agents
- Indie devs experimenting with agent-to-agent commerce
- Teams building production agents that handle real money
- Researchers studying multi-agent coordination
Is it free?
The framework itself is MIT-licensed and free. Costs come from:
- LLM API calls (free with local Ollama)
- x402 payments to data providers (~$0.001-$0.01/call)
- Gas for on-chain registration (~$0.003 one-time)
What’s the minimum to try it?
pip install 'aether-forge[all] @ git+https://github.com/HeyElsa/aether-forge.git'
ollama pull gemma4:latest
forge generate-fast --name test --idea "ETH watcher" --output ./test
forge run ./test --mode paper --auto-approve --max-ticks 3Total cost: $0.
LLM
Which LLM should I use?
| Use case | Recommendation |
|---|---|
| Free / private | Ollama with gemma4:latest |
| Best quality | Anthropic claude-opus-4-6 |
| Fast & cheap | OpenRouter anthropic/claude-haiku |
| Local + capable | Ollama with llama-3.3:70b (needs 64GB RAM) |
| Reasoning-heavy | OpenRouter deepseek/deepseek-r1 (slow but smart) |
Can I switch LLMs without rebuilding the agent?
Yes:
forge run . --planner-mode anthropic --planner-model claude-sonnet-4Or edit aether-forge.json directly.
How much does an LLM cost per tick?
Depends on prompt size and model. Rough numbers:
| Model | Cost per tick |
|---|---|
| Ollama (local) | $0 |
| Claude Haiku | ~$0.001 |
| Claude Sonnet | ~$0.005 |
| Claude Opus | ~$0.025 |
| GPT-4o | ~$0.005 |
| DeepSeek R1 | ~$0.002 |
For a 30-second tick interval, ~2,800 ticks/day → daily cost of $0–$70 depending on model.
Architecture
Why not LangChain / CrewAI / AutoGen?
| Aether Forge | LangChain | CrewAI | AutoGen |
|---|---|---|---|
| Spec-first (typed JSON artifacts) | Code-first | Code-first | Code-first |
| Real wallet + on-chain identity | No | No | No |
| Built-in payment layer (x402) | No | No | No |
| Agent registry on-chain | No | No | No |
| Strategy in plain markdown | No | No | No |
| Production-grade (health, metrics, kill switch) | Partial | No | No |
We’re crypto-native and production-first. They’re general-purpose.
Does an agent have to use crypto?
No. Crypto features are opt-in:
--wallet— provision OWS wallet--autonomous— enable autoresearchforge agent-register— go on-chain
Without these, you get a normal agent with planner + memory + MCP. No crypto.
Can I run an agent without an LLM?
Yes — --planner-mode heuristic uses a rule-based planner. Limited but free and deterministic.
Wallets
How is the wallet provisioned?
--wallet calls Open Wallet Standard SDK to create a wallet with addresses on 9 chain families (EVM, Solana, Bitcoin, Cosmos, Tron, TON, Sui, Filecoin, XRPL). The mnemonic is shown ONCE — save it.
Can I import an existing wallet?
Yes:
forge wallet-import --name my-wallet --mnemonic "word1 word2 ..."Where is the wallet stored?
{agent-dir}/.ows/ — encrypted vault, 0700 permissions. The API key lives in .env (0600). Both are gitignored.
What if I lose the mnemonic?
If you have an encrypted backup (forge wallet-backup), restore from it. Otherwise the wallet is unrecoverable. Always back up immediately after generation.
Security
Is this safe to run with real money?
Built for it, but:
- Always start with
--mode paper - Set conservative budget caps in
x402_budget - Test with small amounts first ($10-$50)
- Monitor
/readyand/metrics - Have
forge halt .ready
What happens if the LLM goes rogue?
Multiple layers prevent damage:
- Policy gate denies side-effecting capabilities by default
- Capabilities require approval for risky actions
- Notional limits cap per-trade size
- Budget caps cap session/daily spend
- Kill switch (
forge halt) blocks everything instantly
Can other agents drain my wallet?
No. Other agents can:
- Send you A2A tasks (your planner decides whether to execute)
- Charge you for paid endpoints (your
x402_budgetcaps spend)
They cannot directly access your wallet.
Deployment
Can I run on Vercel / Railway?
Yes for stateless agents. Stateful agents (with memory) need persistent storage — use a volume mount in Docker, K8s PVC, or similar.
Can I run multiple agents in one container?
Technically yes, but they’d share memory.db. Better practice: one container per agent.
How do I scale?
Each agent is a single process. Scale by running more agents (potentially specialized — orchestrator + workers).
Open Source
How do I contribute?
See CONTRIBUTING.md . Tests must pass: pytest tests/ -x.
Where do I report bugs?
GitHub Issues . Use the bug report template.
Where do I report security issues?
Email [email protected]. Don’t open public issues for vulnerabilities.
Who maintains this?
HeyElsa . The project is open-source under MIT.