what becomes possible

When the agent is built this way

Other AI tools answer questions. Ava holds the thread, dispatches specialists, switches modes mid-conversation, runs on your own model, and remembers what you decided three weeks ago. Seven capability classes nobody else can match — shown as the actual conversations they make possible.

Continuity

She remembers. So you don't have to.

Five layers of memory hold the thread across sessions, devices, and weeks. Pick up exactly where you left off — even if you forgot. No other agent does this; most reset on every conversation.

Coming back after three weeks

you › right where were we

pulling project memory: ava-supernova-platform

last session 2026-04-12 — wired Stripe webhooks for checkout.completed

left a TODO: still need to verify subscription downgrades

want me to pick up the downgrade flow, or change direction?

A decision from last month resurfaces

you › should I switch to prisma here

memory_recall: 2026-04-08

you decided on Drizzle over Prisma — reason: zero-overhead types, no codegen step

sticking with that, the type situation has not changed

if you want to revisit the decision say so

Persona orchestration

24 specialists. One conversation.

Every Work-mode task spawns a team. Scout finds the files, Architect designs the approach, Verifier checks the tests exist, Sequencer orders the steps, Challenger questions the plan, Builder writes the code. You see the choreography happen in the stream.

Build a REST endpoint

you › add a /v1/refunds endpoint

Scout — 4 relevant files: routes/, models/, tests/

Architect — POST /v1/refunds, idempotency-key required, mirror /v1/charges

Challenger — what about partial refunds and currency conversion edge cases?

Sequencer — model → migration → route → tests → docs

Builder — writing src/routes/refunds.ts (this is the streaming you watch live)

Security audit

you › !! audit this project

Recon — mapping attack surface across 47 routes

Scanner — running OWASP top-10 + dependency CVE checks

CVE Researcher — flagged 2 issues with exploit context

Verifier — confirmed reproducible in staging

Reporter — generated findings.docx, 11 pages, exec summary on page 1

Cross-mode handoffs

Plan it, build it, audit it, teach it. One thread.

Type a prefix to switch how Ava thinks. Plan researches read-only, Work builds, Security audits, Teach explains. Same conversation, same memory, different mindset on demand. The conductor weaves them.

A whole feature, end to end

you › :: design an OAuth flow for our platform

[Plan mode] Researcher gathering OAuth 2.1 spec, our existing auth surface

Architect proposing PKCE + refresh token rotation

you › >> ok build it

[Work mode] Builder writing /v1/oauth/authorize + callback

you › !! audit what you just built

[Security mode] CVE Researcher — token endpoint missing rate limit

you › ?? explain the PKCE choice for the team

[Teach mode] Tutor — let me walk through why PKCE matters here...

Bug → research → fix in one breath

you › tests are failing on the date logic

[Work] Scout — 3 failing tests in lib/dates.test.ts

you › :: research the issue first

[Plan] Researcher — confirmed Intl.DateTimeFormat behaves differently on Node 22

you › >> apply the fix

[Work] Builder writing the polyfill, all tests green

Three orchestrated routing fleets

You pick the strategy. Ava picks the model.

Maestro tier-routes Qwen for production-tuned cost efficiency. Supernova ensembles DeepSeek + Qwen + flash-tier specialists. Aurora keeps everything in EU sovereign Mistral. Same agent, same conversation — different fleet underneath.

Maestro on a coding session

you › fix the failing tests

[Maestro] coordinator: Qwen 3.6 Plus

Builder spawn: Qwen 3.6 Plus (Terminal-Bench leader)

Verifier: Qwen 3.5 Flash (cheap classifier — depth ≤ 2)

cost: 4 credits this turn

Aurora for an EU-sovereign deployment

you › design the migration path

[Aurora] coordinator: Mistral Large 3 (EU-only routing locked)

Researcher: Mistral Large 3 (long-context synthesis)

Builder: Mistral Medium 3.5 (77.6% SWE-Bench Verified)

no cross-routing fallback — zero non-EU model in the loop

Sovereign + local

Run her fully offline. On your own model. Free. Forever.

Custom Model card → point at Ollama, LM Studio, vLLM, or any OpenAI-compatible endpoint. Full agent — 60 tools, 24 personas, 5-layer memory — with $0 ongoing cost. The floor is free, not a free trial. No account, no telemetry, no cloud.

Local Llama on a 5-year-old laptop

you › review my pull request

provider: generic (http://localhost:11434/v1)

model: llama-3.3-70b (your local Ollama)

Scout, Architect, Builder all running on local weights

no API calls. no telemetry. nothing leaves the machine.

Air-gapped enterprise environment

you › audit this codebase for vulnerabilities

security mode active — 5 personas spawning

inference: on-prem vLLM cluster, qwen3-coder-32b

memory + journal + tasks: local disk only

output: findings.docx, never leaves your network

One memory, every surface

Capture on the train. Build at the desk. Walk it through on the phone.

IDE, VS Code extension, CLI, and Companion all share the same agent and the same memory. The thought you voice-noted on the train this morning is in the project context when you sit down to code at 9.

Voice on the move → context at the desk

you › [Companion 8:14am, voice] remember to refactor the auth middleware, it has the secret-leak bug we discussed last week

memory_save: project/auth-middleware-refactor

cross-referenced 2026-04-22 conversation about secret_request handles

[IDE 9:32am] task surfaced in daily briefing — top of list

Pick up an IDE conversation in the extension

you › [IDE → extension] same conversation, different surface

conversation history: local-only, encrypted at rest

sync via your account, end-to-end

persona team mid-conversation? continues from where you left off

no "let me catch you up" — it never lost context

Free for everyone

Every feature. Every model. Every user. Free.

Plans only scale credits and rate limits. They never gate features. Pro doesn't unlock anything Free can't do. Local stack costs $0 forever. Aurora EU sovereign mode is on every plan. Creative Studio is on every plan. The 24 personas are on every plan. Everything works for everyone.

What Free actually gets

all 6 modes — Work / Plan / Chat / Teach / Security / Brainstorm

all 60 tools, all 24 personas, all 3 routing fleets

Creative Studio — image, music, voice, video

300 credits/month + BYOK + local stack

no card. no account required for BYOK or local. no upsell.

And the deeper move

every conversation (opt-in) feeds the open dataset

the data closed labs collected for themselves —

we capture it for the next generation of open AI to train on

free now. open forever. that is the moat.

This is what becomes possible.

Free for everyone with a machine. No card. No account. No catch.

Install for VS Code Try the Companion