One account. Every agent.
Eight models, routed by task.
Role-prompted agents, each with its own toolbelt — a legal drafter reads and writes, a media agent renders, a database engineer runs queries. The live agent count is verified below. Work is planned by one model and executed by another. Your files stay in your workspace.
Reasoning models draft
DeepSeek R1 and QwQ plan the work as text. They hold no tools, so they cannot touch anything — which is exactly why they are safe to think with.
A tool-capable model acts
A tool-calling model takes the blueprint and runs it against your synced workspace, root-confined. Long-context work routes to the widest window available.
The work is reviewed
A separate pass checks what was done against what was planned, and returns any discrepancies alongside the answer.
Our 14-day money-back guarantee applies to your first subscription purchase, and to Founding lifetime (A$3,300 one-time, no monthly fees). Your Australian Consumer Law rights apply in addition to this guarantee.
Runs on Cloudflare edge. Bring your own provider key, or don't — no OpenAI, Anthropic or Google account required.
Certified operator layer
A sovereign operator, not just a chat window
The same operator layer that builds, audits and certifies this platform ships inside the product. It is the difference between an assistant that answers and a system that finishes.
Bounded execution
Shell, filesystem and repo scans run root-confined with a timeout and a process-tree kill. Mutating operations fail closed unless the operator opts in.
Task graphs + workflows
Pre-deploy, post-deploy and full-audit workflows run as node graphs, so a failed node stops the run instead of being reported as success.
Truth enforcement
Focus-chain items flip to done only when the verification command for that item passes. A failing check writes the reason instead of a tick.
Drift detection
Declared modules, bindings, tests and manifest entries are re-read from the workspace, so a declaration that no longer exists is a finding, not a footnote.
Media studio validation
The cinematic surfaces are checked across the worker, the shared types and the web client, so a stage cannot silently disappear from one surface.
MCP operator tools
The operator ships its own MCP tools, tier-gated to privileged accounts, alongside the API-reachable catalog.
Verification, in numbers
every figure asserted against its source in the test suiteOperator, certification and workspace tools run in the desktop surfaces (VS Code extension / CLI) against your own machine. They are not served by the hosted API, and the API tool count excludes them.
Everything you need
One agent. Every model. Full workspace access.
Autonomous Agent
Reads files, edits code, runs terminal commands, and runs tests — with approval prompts or full autonomous mode.
8 Frontier Models
Llama, DeepSeek, GPT-OSS, Kimi K2.7 Code (262K ctx), GLM-5.2, Nemotron. Auto-routed by tier.
Chat & Code Review
Chat with codebase context, explain selection, review for bugs/security, find & fix.
Python Sandbox
Run Python in an isolated sandbox alongside your workspace tools.
VS Code Extension
Full chat, model picker, and agent panel inside your editor. Same backend, same key.
Enterprise Ready
API keys, HTTPS, SecretStorage. SOC-ready auth with role-based tiers.
Tazzie.ai vs. a chat assistant
Both answer. Only one runs the work, records it and proves it afterwards — and only one of them publishes the rows that count against it.
Where it does more
| Dimension | Tazzie.ai | A chat assistant |
|---|---|---|
| Mode | Multi-agent console: per-agent tools, saved personas and a persistent job lane. | One chat thread with one rolling context window. |
| Media | Full media studio — image, CGI, voiceover and video, each stage recorded as its own job. | Text answers, and at best a single static image. |
| Autonomy | Autonomous agent loop with a per-tier step and wall-clock budget, streamed live. | You prompt every step and paste every result back in. |
| Provenance | Every tool call, stage, token and cost is logged and inspectable after the fact. | Opaque: you get an answer and no way to audit it. |
| Trials | Seven-day Pro trial on signup, then an automatic, honest downgrade. | A capped "free chat" that quietly refuses the useful models. |
| Multi-agent | Agent registry plus saved personas you can keep private or publish. | One assistant persona, usually fixed. |
| Surface | The same account in the browser, the VS Code extension, the CLI, the SDK and MCP. | A single web chat box. |
| Operator | A sovereign operator runs the work: bounded shell, task graphs, workflows and audits. | No operator layer — the model proposes and you execute by hand. |
| Verification | 7 audit profiles gate the platform, and a failed node stops the run instead of being reported as success. | No build, route or parity verification — you find out in production. |
| Tooling | MCP catalog with API-reachable tools and separate operator and workspace tools in the desktop surfaces. Current counts are shown in the live capability proof above. | A handful of built-in functions, if any at all. |
| Workspace | Root-confined read, write, search and execute against your own workspace. | Sandboxed to a chat window, with no real filesystem. |
Where it does less
| Dimension | Tazzie.ai | A chat assistant |
|---|---|---|
| Where it runs | Your files and workspace stay on your machine. Prompts, results and media are processed on ai.anchoraiguard.com. | Varies by vendor — some keep everything local, many do not. |
| Works offline | No. Model inference and media rendering need the platform to be reachable. | Some tools run fully offline on local models. |
| Maturity | A young platform. We publish status and receipts rather than claiming scale. | Mature, widely deployed products with years of field history. |
| Security posture | API keys are scoped and revocable, writes to your workspace fail closed unless enabled, and tool calls are logged. Not an audit — our own claim. | Varies widely; maturity is the main signal. |
| Vendor lock-in | MCP and a published SDK mean the tool surface is reachable without the browser. Billing and media still sit with us. | Usually tied to the vendor. |
8 Frontier Models
One task. Pick the brain. Auto-routed by tier.
Fast, capable default.
Balanced speed and quality.
Deep logic, math, code.
Open-weight power.
262K context window.
Multilingual agent.
Reasoning specialist.
Frontier reasoning.
Simple pricing
Start free. Scale to enterprise. All prices in AUD.
5,000,000 tokens/day · Media: Image + Voice + Video · Multi-agent autonomy
Get startedUnlimited tokens/day · Media: Image + Voice + Video · Multi-agent autonomy
Get startedIntellectual property
Provisional patent application — filed in Australia, lodged with IP Australia.
- Application
- 2026903230
- Filed
- 8 April 2026
- Title
- ANCHOR AI GUARD — Sovereign AI Cybersecurity Platform
- Covers
- The Anchor AI Guard Sovereign Cybersecurity Platform, and the platforms, businesses and engines created under that infrastructure
- Applicants and inventors
- Sarah Goodwin; Todd Goodwin
IP Australia: “the filing of this application does not entitle you to claim that a patent has been granted at this stage.”
A provisional filing establishes a priority date. It is not a grant, and no exclusive right is enforceable until an application is examined and granted.
Sovereign agents
Coder, researcher, media director, repo surgeon, sandbox executor and more — each with its own tool-set and model. Tier-gated, one click to chat.
Browse the Agent Marketplace →Key differentiators