Your AI agents, your rules
Stop depending on bloated frameworks you don't understand. Built from scratch in Python, so you don't have to. A clean, powerful foundation you fully own and control.
What you get
Everything to build agent systems
From idea to working agent — without stitching together 10 different libraries.
Single & Multi-Agent
Run one agent or orchestrate a team. Agents delegate tasks, share context, and collaborate automatically.
Any Model, One Interface
Use any model from any provider in the same system. Swap providers per agent without changing code.
Secure by Default
3-zone sandbox, command guards, and code scanning built in. Agents are contained — no accidental system access.
How it works
Every agent follows the same reasoning loop
No hidden magic. Every Veska agent runs a transparent Think → Decide → Act → Observe cycle until the task is done.
Think
Analyze the task, context, and available tools
Decide
Choose to use a tool, delegate, or respond directly
Act
Execute the chosen action — run tools, call APIs
Observe
Process the result and decide if more work is needed
Why Veska
Build it yourself vs. Veska
You could wire together multiple libraries and SDKs yourself. Or start with everything already working.
What makes Veska different
3-Level Error Recovery
Other frameworks crash or retry blindly. Veska escalates intelligently, from simple retries to agents collaborating to solve the problem together.
Auto-Retry
Tool timeout, API error, or temporary failure. Veska retries automatically with backoff. No developer code needed.
API rate limited? Retried in 2 seconds. Done.
Agent Self-Fix
Code error, wrong approach, or permission issue. The agent gets the full error context and tries a completely different strategy.
File not found? Agent searches for the correct path and retries with it.
Discussion Room
Cross-agent problem — one agent's output breaks another's work. A Discussion Room opens where agents talk to each other, diagnose the root cause, and agree on a fix.
Backend changed the API format. Frontend agent broke. They discuss and align.
Discussion Room in Action
I changed the API response format. Users now have a "profile" nested object instead of flat fields.
That broke my user dashboard. I was reading user.name but now it's user.profile.name.
I see the issue. Let me document the new schema. All user fields are now under profile.*
Got it. Updating all references to use the new path. Fix applied.
No other framework does this. When agents break each other's work, they fix it themselves.
Features
Everything an agent needs
Production-ready capabilities. Not demos.
Streaming
Real-time token streaming with async generators
Structured Output
Pydantic models with auto-retry on validation
Delegation
Agents delegate to specialists with depth guards
Memory
File, SQLite, ChromaDB, or pgvector backends
Multi-Modal
Images, PDFs, audio — auto-detect or typed objects
Human-in-the-Loop
Agents pause and ask users mid-task when needed
Sessions
Persist conversations per user across restarts
Caching
TTL-based response caching for speed and cost
Philosophy
Why I built Veska
Every existing framework felt like a black box. Too many abstractions, too many hidden decisions, too many dependencies. I wanted a framework where you understand every line and control every behavior.
Read the documentationNo black boxes
Every component is readable Python. No hidden abstractions, no magic methods.
General purpose
Not locked to chatbots or code gen. Research, automation, content, analysis — you decide.
Production ready
Error recovery, security sandbox, cost tracking, and crash recovery built in.
Truly open source
Clone it, fork it, modify it. No vendor lock-in, no paid tiers, no telemetry.
Start building with Veska
Clone the repo, create your first agent, and ship something real.