██████╗ ███████╗██████╗ ███████╗██████╗ ████████╗
██╔══██╗██╔════╝██╔══██╗██╔════╝██╔══██╗╚══██╔══╝
██████╔╝█████╗  ██████╔╝███████╗██████╔╝   ██║
██╔═══╝ ██╔══╝  ██╔══██╗╚════██║██╔═══╝    ██║
██║     ███████╗██║  ██║███████║██║        ██║
╚═╝     ╚══════╝╚═╝  ╚═╝╚══════╝╚═╝        ╚═╝

Your Terminal's Window to the AI World

The keyboard hums, the screen aglow, AI’s wisdom, a steady flow. Will robots take over, it’s quite the fright, Or just provide insights, day and night? We ponder and chat, with code as our guide, Is AI our helper or our human pride?

Your Terminal’s Window to the AI World

Perspt is a high-performance terminal-based LLM interface that serves two purposes: a simple CLI for testing and comparing LLM services across supported providers, and an experimental implementation of the SRBN (Stabilized Recursive Barrier Network) engine inspired by the three-paper Stability is All You Need series. The SRBN agent plans multi-file projects as directed acyclic graphs, verifies each node with real LSP diagnostics and tests, and commits only when Lyapunov energy converges. The theoretical framework is mature; the implementation is under active development.

Quick Start

Install and chat in 5 minutes.

Quick Start

Agent Mode

Autonomous multi-file coding with the experimental SRBN engine.

Agent Mode Tutorial

Dashboard

Monitor agent execution in real-time via browser.

Dashboard

Architecture

Fourteen-crate workspace design.

Architecture

Key Features¶

SRBN Agent	Experimental autonomous multi-file coding under the PSP-9 governed tool loop, with acceptance measured by Lyapunov energy from the SRBN paper series
PSP-10 Mechanisms	Typed prompt programs with ledgered provenance, resident-context paging via `context_recall`, bounded search forests, multi-node work-graph dispatch behind a single integration gate, and crash resume
Multi-Provider	OpenAI, Anthropic, Google Gemini, Groq, Cohere, XAI, DeepSeek, Ollama
LSP Sensors	Real-time type checking via `rust-analyzer`, `ty`, `pyright`, `typescript-language-server`, `gopls`
Test Runner	pytest integration with weighted V_log energy
Per-Role Models	Route models with `--model`/`--actuator-model`/`--explorer-model`/`--adjudicator-model`, or per role (architect, actuator, verifier, speculator, adjudicator) via the `[models]` table
Token Budget	Cost control with usage monitoring and per-request limits
Beautiful TUI	Ratatui-based with diff viewer, task tree, dashboard, and review modal
Security	Starlark policy engine with command sanitization and workspace-bound enforcement
MCP 2026-07-28 Client	Official-SDK stdio/Streamable HTTP support for tools, context primitives, client requests, subscriptions and tasks, with local admission and `/mcp` diagnostics
Merkle Ledger	Cryptographic change tracking with session resume and rollback
Headless Mode	Fully autonomous operation with `--yes` for CI/CD and batch workflows