Architecture for Agentic Apps.
Modern AI apps are tangled — prompts, tools, and orchestration are buried in application logic, untestable until production, opaque when things break, and impossible to hand off to the people who understand the domain.
Declarative Definitions
Define Expert behavior in TOML. Integrate via API. Deploy definitions and app code independently.
Toolkit for Developers
Run locally with production-identical isolation. Inspect every tool call, delegation, and checkpoint in detail.
Domain-Native Tuning
Domain experts tune behavior in TOML. Developers own integration. Each side ships on its own.
Deterministic Execution
Every run is a replayable event stream with step-level checkpoints. Diff to find drift.
Optimized for Sandbox
Each Expert runs in its own isolated context. Workspace boundaries, environment sandboxing, tool whitelisting.
Cross-Job Analytics
Success rates, token usage, errors, and tool utilization across jobs. CLI and Studio dashboards.
Define. Experts in TOML, not code.
Expert definitions in perstack.toml are written by domain experts using natural language. Developers focus on integration, not prompt engineering.
Declarative TOML
Define Experts with instruction, skills, and delegates. No classes, no boilerplate, no framework to learn.
MCP-native skills
Connect to any MCP server as a skill. GitHub, Slack, databases — your Expert uses real tools, not wrappers.
Multi-provider
Anthropic, OpenAI, Google, DeepSeek, Ollama, Azure, Bedrock, Vertex. Switch providers with one config change.
Get a working Expert in one command.
Describe what you need. create-expert generates the definition, tests it against real scenarios, iterates until behavior stabilizes, and reports what it can do.
$ npx create-expert "Create a code-reviewer that delegates to a style-checker"The result is a perstack.toml you can run immediately — or customize further:
[provider]
providerName = "anthropic"
model = "claude-sonnet-4-5"
maxSteps = 100
[experts."code-reviewer"]
description = "Reviews pull requests for bugs and style issues"
instruction = """
You are a senior code reviewer. For each PR:
1. Check for bugs, security issues, and performance problems
2. Verify test coverage
3. Delegate to style-checker for formatting
"""
delegates = ["@code-reviewer/style-checker"]
[experts."@code-reviewer/style-checker"]
description = "Checks code style and formatting"
instruction = "Check code against project style guidelines."
[experts."code-reviewer".skills.github]
type = "mcpStdioSkill"
command = "npx"
packageName = "@anthropic/mcp-github"Run. Same Behavior, Everywhere.
Open-source runtime. Event-derived execution. Deterministic checkpoints.
$ npx perstack start code-reviewer "Review PR #42 for security issues"// observable 24 activity types, real-time
Every tool call, every delegation, every LLM interaction—streamed as structured events. Debug anything.
// isolated Ephemeral VM per job
Each Expert runs in its own isolated context — workspace boundaries, environment sandboxing, and tool whitelisting. No shared state between runs.
// reproducible Lockfile for production
Pin skill versions, provider configs, and Expert definitions. Event-derived execution and step-level checkpoints maintain reproducible behavior.
How Perstack Compares
Build your first Expert in 5 minutes.
Create. Execute. Integrate.