Consistent CEO

Inspiration

Most AI tools are temporary. You ask a question, you get an answer, and then the context fades. The real state of what you're building scatters back across chat threads, folders, repos, notes, emails and decisions in your head.

I wanted a private, always-on assistant that remembers what I'm building and actually follows through, running on my own computer, with my data under my control. That became Consistent CEO.

What it does

Consistent CEO is an open-source (MIT), self-hosted AI workspace for living entities: a business, a project, a codebase, a goal, or you.

• Entities with persistent memory. Each entity holds its files, folders, repos, notes, decisions and history, with semantic recall, so you never re-explain context.
• AI directors. Specialised agents per entity, each with its own lessons and scheduled heartbeats, so work keeps moving while you're away.
• Missions with proof. You set a definition of done. The system loops develop → verify → develop, judged on real file changes and test results (e.g. a git diff or a passing test). Runs that claim success without changing anything are flagged.
• Safe by default. Unattended code and file work is confined to a folder you choose. Browser work uses a separate profile. Anything outward-facing (sending messages, network access) waits for your approval.
• Private by design. Runs locally with no project-run account or server. Encrypted vault unlocked by password, passkey or biometrics. A phone companion pairs by QR code over your own network for quick capture, approvals and heartbeat notifications.

How I built it

• Core: one Rust process (consistent-world) serves the HTTP API and the SvelteKit dashboard and runs all background loops (missions, directors, heartbeats). Crates: gateway, agents, llm, memory, tools, plugins, credentials, architect, relay.
• Apps: Tauri desktop shell (macOS and Windows) and an Expo mobile companion.
• Memory: embedded SurrealDB with vector search for sessions, memories, embeddings and semantic recall.
• Models: the llm crate is provider-agnostic: Claude Code and Codex CLIs, ChatGPT subscription, local models via Ollama, or API keys for OpenAI, Anthropic, Google or any OpenAI-compatible server. Because Nebius Token Factory is OpenAI-compatible, the same layer is the path for running NVIDIA Nemotron models.
• Safety: sandboxed WASM plugins, OS keychain plus encrypted vault, approval gates for outward actions, and an architect crate that shadow-verifies answers.
• Packaging: desktop installers, an install script with a background service, and Docker.

Challenges I ran into

• Making "done" verifiable rather than self-reported, so the loop checks evidence (diffs, tests) instead of trusting the model's claim.
• Keeping autonomous loops safe on a real machine: folder confinement, approvals and clean separation of browser profiles.
• Keeping long-term memory useful rather than noisy as entities grow over weeks and months.

Accomplishments that I'm proud of

• A working, desktop-first personal AI that remembers, verifies its own work, and asks before acting outside its sandbox.
• Released fully open source under MIT, so anyone can run, audit and extend it.

What I learned

Persistent context plus verification matters more than raw model size. Keeping the model layer provider-agnostic means the same assistant can run on cloud APIs, open models or fully local models without changing how it remembers and verifies work.

What's next for Consistent CEO

• Smoother onboarding and one-click provider setup (including Token Factory).
• More directors, skills and plugins.
• Optional Nebius Serverless for always-on cloud heartbeats while keeping data local-first.

Built With

• docker
• expo.io
• ollama
• rust
• surrealdb
• svelte
• sveltekit
• tauri
• typescript
• webassembly