Reconstruct agent architecture. Track runtime executions. Debug nodes in isolation.
Cohzen (cz) statically scans LangGraph AST and captures zero-code execution telemetry. Run cz runs to inspect timeline waterfalls, and cz node run to debug handlers directly in isolation.
AI isn't a model anymore.
It's a system.
Models, prompts, context, tools, memory, and agents work together.
Change one—and the behavior can change everywhere.
Hover over any node to inspect real-time latency, token budgets, and runtime boundaries.
AI agents fail differently.
Traditional software gives you a stack trace.
Agents give you a trail of decisions.
The error isn’t where it started.
You shouldn't have to
reconstruct the story.
Cohzen keeps the whole execution together.
The control layer for your AI system.
Observe it. Replay it. Evaluate it. Prevent it.
See your actual agent graph architecture.
One layer. One execution. One source of truth.
From production failure to verified fix.
Cohzen turns a difficult production failure into an execution you can inspect, replay, change, evaluate, and verify.
Reconstruct your agent architecture from AST.
Run cz scan in any repository. Zero code rewrites, zero package imports, zero API keys. Cohzen builds a complete graph model in milliseconds.
Audit telemetry coverage and eliminate blind spots.
cz audit cross-references package manifests, environment configurations, and function-level decorators to show which agent nodes lack tracing before shipping.
Inspect in your browser or generate Mermaid flowcharts.
Explore your agent graph topology with pan/zoom, handler line spans, and tool schemas locally, or export Mermaid flowcharts for PR documentation.
Initialize zero-code execution tracking with cz init.
Run cz init once in your repository. In v0.3.4, it renders a slim topology report, installs a zero-overhead virtualenv .pth startup hook, and records LangGraph executions into local SQLite with zero network egress.
Emit the immutable CohzenManifest schema.
cz scan --json emits a canonical, machine-readable JSON contract with stable IDs and precise source locations for compliance and reporting.
Enforce architectural gates in CI/CD.
Add Cohzen to your GitHub Actions workflow. Block PRs that introduce uninstrumented LLM caller nodes, missing tracing keys, or unmapped conditional routes.
Every failure makes your system smarter.
The execution that broke in production doesn't disappear. It becomes something you can understand, replay, evaluate, and turn into a guardrail.
Live agent serving real user requests.
Agents process multi-step workflows, orchestrate tools, and mutate context in production. Every execution boundary is instrumented with zero latency overhead.
Your production data shouldn't just tell you what went wrong.
It should make your AI better.
Every edge case crystallized into a test. Every regression prevented before merge.
Scan and audit in your terminal first.
Cohzen v0.3.4 pairs static LangGraph AST scanning with zero-egress runtime capture, virtualenv startup hooks, and CI/CD audit gates. Map graphs with cz scan, initialize transparent capture with cz init, trace executions with cz runs, and inspect detailed spans with cz run.
cz scan .Reconstruct AST in milliseconds
Statically reconstructs LangGraph workflows directly from Python AST without executing any runtime code or importing unverified packages.
Deterministic static analysis, zero runtime execution.
Cohzen reconstructs what can be determined statically from the codebase directly from Python AST. Zero imports. Zero API keys. Zero side effects.
100% Static AST Reconstruction
Cohzen traverses Python source code using Python’s native ast parser. It never imports untested code, never executes closures, and never requires active secrets like OPENAI_API_KEY.
The Immutable System Manifest (v0.1)
The canonical CohzenManifest represents graphs, nodes, tools, and models with exact source line spans, resolution states, and completeness metadata decoupled from presentation formats.
Keep the tools you already use.
Cohzen is designed to be composable. Keep your agent framework. Keep your observability platform. Keep your evaluation tools. Cohzen connects the debugging loop between them.
The framework may change. The debugging problem doesn't.
Not another dashboard.
Cohzen isn't trying to become the place where every metric, trace, and log lives. There are already excellent tools for that. Cohzen focuses on the gap between:
“Something failed in production.”
“I understand why it failed. I've reproduced it. I've fixed it. And I've made sure it won't happen again.”
That's the loop Cohzen is built to close.
Your AI infrastructure should be inspectable.
Cohzen is open source and self-hostable. Inspect the code. Understand how executions are captured. Run the infrastructure where your systems run.
Building the reliability layer for AI systems.
100% static LangGraph AST scanner (cz scan), telemetry audit (cz audit), CohzenManifest v0.1.0, Mermaid export, HTML UI (cz view).
Observability gap auditing (cz audit), standalone interactive HTML visualizer (cz view), and initial runtime telemetry models.
Virtualenv .pth hooks, slim cz init topology output, lifecycle controls (cz on/off/status/deinit), one-off capture, SQLite secret redaction, and CI audit gates (--min-coverage, --fail-on-empty).
Static architecture & runtime instrumentation for LlamaIndex Workflows, CrewAI, AutoGen, and multi-agent coordination.
Deterministic execution replay, failure boundary isolation, automated regression generation, and centralized manifest registry.
Frequently asked questions.
Your agent already ran.
Now understand why.
Open source · Developer first · Built for AI systems