From an idea to an agent. With knowledge that stays with you.
A local agent foundry, research workspace and knowledge system powered by NeuralDeep.
Get started · Installation · Example agent · Search & research
Pritha NeuralDeep turns an idea, project or workflow into a specialist-agent project you can inspect, test and improve. It brings together a complete local Control Center, a curated engineering knowledge base and NeuralDeep model access.
This is an independent distribution of Pritha. It includes the platform and its reusable knowledge, with a fresh workspace for your own tasks. Your API keys, conversations and agent data stay out of Git.
NeuralDeep is our preferred provider for this edition: convenient, cost-effective, and a great fit for everyday agent work. One provider connection brings together model execution, Voice Control, and web search. That means less integration work and a straightforward path from installing Pritha to using its full workspace. Brief Desk ND can reuse the same NeuralDeep credential, making it easy to start working with a real agent straight away.
The installation and agent-building experiences below describe what this feels like in practice.
- Download the repository as a ZIP and extract it.
- Open the extracted folder in Cursor, Codex, Claude Code, or another coding assistant that can read files and run local commands.
- Ask:
Read START_HERE.md, then set up and start Pritha NeuralDeep.
The assistant installs dependencies, rebuilds local memory, sets up an isolated instance and opens the local Control Center. Enter your own NeuralDeep API key in Settings → NeuralDeep. Do not paste keys into the assistant conversation.
Need a key? Register with NeuralDeep, copy the key from your dashboard, and connect it in Pritha Settings. Quick account and API-key guide — including provider advantages and access to the default Kimi model.
Prefer a terminal?
node scripts/bootstrap.mjs prepare --profile neuraldeep
node scripts/bootstrap.mjs start --profile neuraldeepRequirements: Node.js 24+, Python 3.10+, Git, SQLite CLI and Codex CLI. The installation assistant can be any suitable coding tool; Codex CLI is still used internally to execute agent tasks through the NeuralDeep adapter.
macOS is the primary tested platform. Linux and Windows are not blocked, but installation, credential setup and service management are self-managed and have not received full end-to-end verification. See installation details.
| Surface | What you can do |
|---|---|
| Agents | Design, inspect, test and evolve specialist agents using contracts and outcome specifications. |
| Task Chat | Work on tasks with persistent local history, attachments and execution status. |
| Voice Control | Talk to Pritha through NeuralDeep speech and dialogue, and hand work to the task executor. |
| Search & Research | Use shared web search and source reading in Chat and Voice, or run a bounded research task. |
| Knowledge | Reuse curated standards, decisions, workflows, templates and engineering lessons. |
| Settings | Connect NeuralDeep, choose models, inspect usage and configure optional integrations. |
The distribution starts with Kimi for tasks, NeuralDeep Voice, and NeuralDeep Search in Auto mode. Search is enabled for Chat, Voice and Research; child-agent access still requires an explicit allowlist. Automatic SearXNG fallback is off. These defaults match the accepted local configuration; search relevance and source freshness still need review. Auto is configurable in Settings and makes metered provider calls when a task needs external information.
All service links start on localhost. Tailscale, remote access and automatic service startup are separate opt-in setup steps.
One complete example agent is included:
topic → sources → draft → edit → approve → Telegram
Open its card in Agents, click Start, then open its local interface. Its offline demo works without API keys. Settings lets you use the parent Pritha NeuralDeep credential or an independent key, and optionally connect a Telegram bot and channel. Telegram is not required to create and edit briefs. Nothing is posted automatically.
Brief Desk guide · Connect Telegram
No other live agents are bundled. Task Chat, Voice, Drafts and Brief Desk history start empty; historical engineering lessons remain available as shared knowledge.
- Code and knowledge: platform code and curated knowledge in this repository. Raw intake notes stay local to each instance and are not published.
- Local state: a separate instance directory for sessions, settings, queues and generated indexes.
- Your agents: a separate directory belonging to this instance; unrelated sibling folders are not imported.
- Memory: rebuilt from curated Markdown. Pre-existing conversation databases are never shipped.
- Credentials: macOS Keychain for Pritha; server environment for manual setups; local ignored configuration for a Brief Desk override.
Model, speech and search requests go to NeuralDeep. Optional Telegram publication sends the approved brief to your selected destination. Local-first does not mean inference is offline.
- Get a NeuralDeep account and API key
- Installation and first run
- NeuralDeep execution
- Search and bounded research
- Architecture · Memory · Engineering rules
- Release and updates
- Contributing · Security
- Model-written session reviews (qualitative, supplied by the operator)
The public getting-started documentation is in English. The inherited engineering knowledge base and parts of the application retain their original Russian content.
MIT. Based on NumericalArt/Pritha, with a separate NeuralDeep runtime, release process and bundled example agent.