STOA Agent ⁂

Six personas in one chamber, running on your machine. Council-mode debate · Bring-your-own-key · No subscription, no lock-in.

STOA is a Socratic chamber: six named personas — each with its own role, system prompt, and reasoning style — debate every non-trivial task in parallel, then a seventh dispatcher composes a verdict. STOA itself is free and open-source forever; you bind your own LLM API key (we recommend DeepSeek's free tier — 2-minute signup, no card) and STOA never sees, brokers, or proxies your key. Want one provider per persona (e.g. Sokrates → Anthropic, Veritas → Google, Drax → xAI)? Bind a different key per seat in ~/.stoa/cli-config.yaml.

This is a fork of NousResearch / hermes-agent v0.14.0 (MIT). The runtime, gateway, sandboxes, memory store, skill format, and provider plugin layer are inherited from upstream — see ATTRIBUTION.md for the full lineage. What STOA adds on top is the chamber: persona orchestration, an optional on-chain attestation preview, and a 6-agent skill-publication audit gate.

Install

macOS · Linux · WSL2 · Termux

curl -fsSL https://stoax.xyz/install.sh | sh

Windows · PowerShell

iex (irm https://stoax.xyz/install.ps1)

PyPI

pip install stoa-agent
stoa setup

⚠️ Install stoa-agent, not stoa. The bare stoa name on PyPI is an unrelated third-party package — pip install stoa does not install this project. Always use pip install stoa-agent (or the official installer scripts above).

Homebrew

brew tap stoagent/stoa
brew install stoa-agent

Direct from source

git clone https://github.com/STOAGENT/stoa-agent
cd stoa-agent
uv venv && uv pip install -e .
stoa setup

What STOA adds

1. Council mode — six personas, one chamber

Each persona has its own system prompt, reasoning style, and tool affinity. They run in parallel against the same task, and a seventh dispatcher composes a verdict by surfacing agreement and dissent rather than collapsing the answer to a single voice.

stoa /council "audit this contract: $(cat MyToken.sol)"
# → 6 personas in parallel
# → Sokrates / Mira / Veritas / Drax / Lyra / Echo each respond
# → Verdict + per-persona dissent + response hash (local)

2. Provider model — bring-your-own-key

STOA never bundles, brokers, or proxies anyone else's API key. There is no STOA cloud and no "free credits from us." Instead the first-run wizard walks you through the cheapest viable setup:

1. Go to https://platform.deepseek.com/api_keys (free signup, no card).
2. Create a key. DeepSeek's free tier covers typical solo use; a heavy session is a few cents.
3. Paste the key. STOA writes it to ~/.stoa/cli-config.yaml on your machine and never sends it anywhere else.

Result: the chamber works end-to-end on your own DeepSeek free tier, with you in full control of the spend. Want one provider per persona (e.g. Sokrates → Anthropic, Veritas → Google, Drax → xAI)? Bind a different key per seat:

# ~/.stoa/cli-config.yaml — produced by `stoa setup`, fully editable
personas:
  sokrates: { provider: deepseek, model: deepseek-reasoner }   # default
  mira:     { provider: deepseek, model: deepseek-chat }       # default
  # …or override per seat with a different key + provider:
  veritas:  { provider: anthropic, model: claude-opus-4-7, api_mode: anthropic }

The persona names (Sokrates / Mira / Veritas / Drax / Lyra / Echo / Hermes) are role identifiers and are decoupled from any single model vendor — see stoa /persona list for the live binding on your machine.

3. Council-audited skill publication

The hardest problem in agent skill ecosystems is supply-chain trust. STOA's answer: no skill publishes without a 6-persona audit + 5-of-6 quorum + a local audit hash. Security, performance, prompt-injection, license, structure, attribution — six different lenses on every new skill.

stoa skill publish ./my-skill
# → 6 personas audit it independently
# → 5-of-6 quorum required
# → audit hash written locally; on-chain stamp behind --attest (preview)

4. On-chain attestation — preview

stoa --attest is currently a preview feature behind a flag.

When enabled, every council verdict optionally writes its response hash to AuditAttestationV2 on Monad mainnet, so months later anyone can verify a STOA agent ran exactly the action it claims it ran. The hashing + persistence are wired; the eth_sendRawTransaction submission and verifier client are under hardening for the next release. Until then, expect --attest to compute the hash, queue the request, and log attestation_preview: pending_submit.

If you have no need for on-chain verifiability, you can ignore --attest entirely — the chamber, the verdict, and the skill audit gate all work locally without it.

Commands

Skills shipped under skills/stoa/

• council-verdict — orchestrate a 6-persona call from inside a skill
• monad-attestation — write a hash to AuditAttestationV2 (preview)
• solidity-audit-pipeline — Slither + Mythril + Echidna + manual review
• erc8004-reputation — read or write agent reputation events
• stoa-skill-publish — the publication audit gate itself
• monad-mev-watchdog — passive on-chain monitor
• solana-anchor-audit — Anchor-program review

Security posture

STOA inherits the same primitive set as its upstream lineage: shell execution, browser automation, plugin marketplace, optional wallet binding. These are powerful tools and require operator literacy — STOA targets the same operator profile as Cursor, Claude Code, and Aider.

• Default-OFF gates (DB encryption, PII/IP redaction, skill ed25519 signature, mandatory sandbox) are being flipped to default-ON via a STOA_SECURITY_PRESET selector in the next release.
• Bug bounty + coordinated disclosure: see SECURITY.md. Do not open public issues for security reports.
• Audit reports are not published to the master tree — coordinated disclosure first. We ship fixes, then summary write-ups.

License

MIT for the STOA Agent codebase. See LICENSE. The upstream MIT license is preserved verbatim; this fork adds the attribution recorded in ATTRIBUTION.md.

Bundled assets carry their own licenses:

• web/public/fonts-terminal/JetBrainsMono-*.woff2 — SIL Open Font License 1.1, see web/public/fonts-terminal/OFL.txt.
• optional-skills/productivity/powerpoint/ — Proprietary, Anthropic. Opt-in only (set STOA_ENABLE_OPTIONAL_SKILLS=1 to discover it). Use is governed by your separate agreement with Anthropic; the file optional-skills/productivity/powerpoint/LICENSE.txt ships the full terms. NOT covered by MIT.
• optional-skills/mlops/inference/obliteratus/ — AGPL-3.0. Opt-in via STOA_ENABLE_REDTEAM=1. AGPL §13 obligations apply if you ship a network-accessible service that incorporates this skill.

Links

• Docs · https://stoax.xyz/cli
• Chamber · https://stoax.xyz
• PyPI · https://pypi.org/project/stoa-agent/
• Source · https://github.com/STOAGENT/stoa-agent
• Upstream lineage · ATTRIBUTION.md