PSI Studio Technical Documentation

Quickstart

Integrate PSI Studio into your CI/CD or developer environment using the official SDK or CLI runner:

npm install -g @piron/psi-cli
# Or in Python:
pip install psi-orchestrator

Initialize a deterministic repository workspace session:

export ANTHROPIC_API_KEY="sk-ant-..."
psi-cli init --engine claude-sonnet-5-5 --mcp git,test_runner,ast_scan

Core Architecture

Traditional AI coding assistants suffer from hallucinations because of ungrounded token generation. PSI Studio implements a 4-phase closed loop:

• Planner (Claude Sonnet 5.5 / Opus 5.5): Ingests the full repository AST and task specifications via a 1,000,000 token context window with prompt caching.
• MCP Tool Dispatcher: Executes deterministic, structured tool calls against local development environments (Git, PostgreSQL, Docker, Linter) over stdio and SSE transport.
• Verifier: Executes automated test suites and compiler passes (`tsc --noEmit`, `pytest`, `cargo test`) in an isolated container.
• Auditor: Commits only verified patches with cryptographic integrity checks.

Claude 5.5 Generation Integration

PSI Studio targets Anthropic's flagship claude-sonnet-5-5 and claude-opus-5-5 models with adaptive thinking and 1M token context windows:

{
  "model": "claude-sonnet-5-5",
  "max_tokens": 16384,
  "thinking": {
    "type": "adaptive",
    "effort": "high"
  },
  "system": [
    {
      "type": "text",
      "text": "You are PSI Studio Lead AI Architect. You solve tasks using deterministic MCP tool calls and verify all code changes against test suites before reporting completion.",
      "cache_control": { "type": "ephemeral" }
    }
  ],
  "tools": [ /* MCP Tool Schemas */ ]
}

High-frequency telemetry routing, AST classification, and lightweight subagents are dispatched to claude-haiku-5-5 for sub-second execution.

Model Context Protocol (MCP) Support

PSI Studio natively conforms to Anthropic's open Model Context Protocol standard. Agents connect to any compliant MCP server to access external systems safely:

• mcp-server-git: Inspect commits, branches, and diffs with AST awareness.
• mcp-server-postgres: Execute query explain plans and generate optimized migrations.
• mcp-server-docker: Spin up isolated test runners and ephemeral databases.

REST API Reference

Trigger Autonomous Workflow: POST /v1/orchestrations

Starts a multi-agent orchestration loop on a targeted repository.

curl -X POST https://api.julienpiron.com/v1/orchestrations \
  -H "Authorization: Bearer psi_sec_..." \
  -H "Content-Type: application/json" \
  -d '{
    "repository": "github.com/org/service",
    "task": "Audit authentication pipeline for CVE-2026-1049 and apply verified fix",
    "model": "claude-sonnet-5-5",
    "effort": "high",
    "max_iterations": 10
  }'