Your AI agents work autonomously — but do you actually know what they're doing? Satya gives agents a task board, permission scope, and an audit trail humans can watch in real time.
Satya enforces a clean separation: the AI agent owns all operations. The human gets a live browser dashboard — no terminal needed.
LangSmith, Langfuse, AgentOps — they all tell you what your agent did. Nobody was asking what your agent is allowed to do.
Your agent has no task scope. It deletes files it wasn't supposed to touch, marks itself done, and charges you $40 in API costs.
You define the scope. Satya enforces it. Every action is logged before it happens. Humans monitor in real time with zero setup.
allowed_actions & forbidden_actionsDrop into any Python agent in seconds. No database, no cloud, no config file needed.
# 1. Install pip install satya-ai # 2. Use in your agent — 3 lines import satya.sdk as satya client = satya.init("my_agent") # start session task = client.create_task("Analyze pricing", "Compare competitors") # create task satya.log("Starting analysis...") # log progress # Agent does the work... client.update_task(task["id"], "Done") # mark done (checker validates) client.flush_logs() # persist via git (optional) # Human opens http://localhost:5000 — sees everything above in real time. # No terminal. No config. Just a browser.
Zero infrastructure. Flat JSON files. Git-compatible. Works with Claude Code, Jules, Codex — any agent with Python access.
Kanban-style columns (To Do / In Progress / Done) with priorities. Agents update it. Humans watch it. No refresh needed.
Per-task allowed_actions and forbidden_actions. The can_do() function blocks everything not explicitly allowed. Default-deny.
Agents can't self-certify. Verify by file existence, test pass, subtask completion, or manual human approval.
Scrape any URL to clean Markdown, stored in a shared knowledge base. Agents and humans both reference the same ground truth.
Timestamped entries from every connected agent, visible in real time on the dashboard. Full audit trail, zero setup.
Automatically detects stale tasks — tasks that haven't moved in too long. Surfaces them on the dashboard with a warning badge.
Drop-in polling runner for Claude Code and compatible agents. Checks for new tasks, executes, reports — fully autonomous.
Everything lives in flat JSON + Markdown files in satya_data/. No database, no cloud service, no migration scripts.
One-click toggle in the sidebar. Full CSS variable swapping. Works great on the projector in your next demo.
Each star signals to the AI community that open-source agent governance matters. Help us hit 200 stars this week.
Star on GitHub — github.com/AnkTechsol/Satya_AI