A local-first agent.
You approve the actions.
ULTRON is a local prototype designed to run on your machine. It reads, plans, and proposes — with an approval gate designed to pause trades, deploys, file writes and mutating API calls for a human decision: Y or N.
How it works
Two tiers. One gate. An immutable hash-chained audit log behind every decision.
Decompose the objective
The planner turns your natural-language goal into a DAG of tool calls. Nothing runs yet.
Read-only tools auto-run
Filesystem reads, web search, code-index queries, market quotes, Drive search — all run in-process with no approval.
Side-effects queue a Proposal
Every write, trade, deploy, or mutating API call pauses and emits a typed Proposal — intent, exact command, rollback plan, risk 1–5.
You decide with Y or N
Approve from CLI, HUD, or API. The gate re-checks the kill switch, runs the tool, and hashes the result into the audit chain.
Command Deck · Sample proposals
This interactive UI contains three sample proposals across risk levels. APPROVE and DENY illustrate the review flow; they do not execute live tools, trades or deployments.
Sample UI only — no live execution. Demo mode uses sample proposals. Source access and local setup are by request. The local daemon entry point is python3 -m core.api.
The five invariants
The engine is designed to require explicit APPROVE for the actions below, with a hash-chained audit log and a kill switch. These controls depend on the implementation, configuration and tools you connect; they are not a guarantee against every failure.
file_write and git push passes through the gate with a rollback plan.ConsentScope. Out-of-scope reads are refused at the memory layer.Local prototype. Hosted options are planned.
ULTRON is a local prototype with source access by request; a public source download is not currently available. Hosted Pro and Team options are planned and waitlist-only: no active paid plans, checkout or trial is available. Features, pricing and launch timing are not finalized.
Local prototype
- All 21 tools + all 5 integrations
- Ollama / Anthropic / Perplexity adapters
- Full audit chain + kill switch
- 30 adversarial tests
- Source access by request
Hosted Pro · Planned
- Local engine capabilities
- Hosted API daemon (Railway-backed)
- Managed Anthropic + Perplexity keys
- Cloud-synced audit log backup
- Priority support + invariant bounty
Team · Planned
- Everything in Pro
- Up to 10 operators
- Shared personas + role-based approvers
- Centralized audit across operators
- SSO + SAML (on request)
Join the hosted-interest waitlist
Leave your email for updates about planned hosted options. Joining the waitlist does not create an account, start a trial or activate a paid plan. No launch date is promised.
Your email goes to Clayton's own CRM — no third-party lists. Hosted availability updates only; no payment is collected here.