Kestrion

A durable-execution-first framework for building production AI agents.

Status: v0.7 — actively developed. Published on PyPI. Core engine, the Agent/@tool decorator API, three LLM providers, a live-verified MCP client and server, a CLI, twenty-three agentic features (vision/multi-modal support, multi-step approval chains, time-boxed approvals, parallel tool calls, sub-agents, multi-agent handoff, memory/context compaction, human-in-the-loop input, enterprise secret management, advanced context window management, swarm routing via supervisor nodes, dynamic tool discovery, visual playground builder, browser automation, Vercel Generative UI, secure code sandboxing, JSON mode / structured outputs, Retrieval-Augmented Generation / RAG, Data Loss Prevention / PII Redaction, Cross-Run Episodic Memory, Computer-Use Desktop Automation, Real-Time Streaming Server, and Multi-Verse Execution), and a DAG-based async scheduler for multi-agent orchestration are built and tested. Horizontally scalable, multi-worker Postgres support and rate-limited scheduler are built.

See Roadmap below.

Why Kestrion

Most agent frameworks are strong at authoring an agent loop. Kestrion is built around a narrower, specific bet: state is never mutated directly — it's derived by folding an immutable log of events. That single decision is what makes the following true by construction, not by careful discipline on the part of whoever writes a given agent:

• Crash recovery is the default. Any process — the original one or a brand new one — can reconstruct a run's exact state from the store and continue it.
• Human-approval gates pause the run itself, not just a function call. A tool marked as requiring approval can't be invoked without it, enforced centrally by the engine.
• Observability comes from the same log everything else does — token counts, cost, and full trace history, not a separate system bolted on after.

Architecture

flowchart TD
    %% Define Styles
    classDef user fill:#eef2ff,stroke:#6366f1,stroke-width:2px,color:#312e81
    classDef agent fill:#f0fdf4,stroke:#22c55e,stroke-width:2px,color:#14532d
    classDef engine fill:#fffbeb,stroke:#f59e0b,stroke-width:2px,color:#78350f
    classDef store fill:#f8fafc,stroke:#cbd5e1,stroke-width:2px,color:#334155
    classDef llm fill:#fdf4ff,stroke:#d946ef,stroke-width:2px,color:#701a75
    classDef tools fill:#ecfdf5,stroke:#10b981,stroke-width:2px,color:#064e3b

    %% Actors & High Level
    User(("🧑‍💻 User / App")):::user
    Agent["🤖 Agent (kestrion.agent.Agent)"]:::agent
    
    User -- "run('Fix the bug...')" --> Agent
    
    %% Core Engine Flow
    subgraph Core ["Core Execution Engine"]
        Engine["⚙️ Engine (State Machine)"]:::engine
        State["📦 AgentState (scratch dict)"]:::engine
        
        Node_Entry("Node: __entry__"):::engine
        Node_LLM("Node: __llm__"):::engine
        Node_Tools("Node: __tools__"):::engine
        
        Engine --> State
        Engine -. "Routes" .-> Node_Entry
        Node_Entry --> Node_LLM
        Node_LLM -- "Requires Tool" --> Node_Tools
        Node_Tools -- "Tool Output" --> Node_LLM
        Node_LLM -- "Final Answer" --> Finish(("End")):::engine
    end
    
    Agent -- "Drives" --> Engine
    
    %% Storage Layer
    subgraph Storage ["Persistence (SQLite)"]
        Store[("💽 SQLiteCheckpointStore")]:::store
        Checkpoints[/"Table: checkpoints"\]:::store
        Events[/"Table: events"\]:::store
        
        Store --> Checkpoints
        Store --> Events
    end
    
    Engine -- "Log Event / Save Checkpoint" --> Store
    Store -- "Fork / Resume / Time-Travel" --> Engine
    
    %% LLM Providers
    subgraph Models ["LLM Provider Abstraction"]
        Provider["🔌 LLMProvider Interface"]:::llm
        Anthropic["AnthropicProvider"]:::llm
        OpenAI["OpenAIProvider (Gemini, etc)"]:::llm
        Ollama["RobustOllamaProvider"]:::llm
        
        Provider -.-> Anthropic
        Provider -.-> OpenAI
        Provider -.-> Ollama
    end
    
    Node_LLM -- "complete()" --> Provider
    
    %% Tools
    subgraph Toolkit ["Tool Execution"]
        LocalTools["Local Functions (@tool)"]:::tools
        Approval["Human-in-the-Loop Gates"]:::tools
        Sandbox["CodeSandboxToolkit"]:::tools
    end
    
    Node_Tools -- "Executes" --> LocalTools
    Node_Tools -- "Wait for Approval" --> Approval
    LocalTools -.-> Sandbox

Install

pip install kestrion[anthropic]   # or [openai], [ollama], [mcp], [browser], or [all]

Each LLM provider and MCP support are optional extras. If you only use Ollama, you never need the anthropic or openai packages installed.

Quickstart

import asyncio
from kestrion.agent.agent import Agent
from kestrion.agent.decorators import tool
from kestrion.llm.anthropic_provider import AnthropicProvider

@tool
def get_cluster_state() -> dict:
    """Read current deployment replica counts."""
    return {"deployment": "checkout-api", "replicas": 2}

@tool(requires_approval=True)
def apply_manifest(yaml: str) -> dict:
    """kubectl apply a manifest against the cluster."""
    # real kubectl call would go here
    return {"applied": True}

async def main():
    agent = Agent(
        provider=AnthropicProvider(model="claude-sonnet-4-6"),
        tools=[get_cluster_state, apply_manifest],
        store="sqlite:///agent_runs.db",
    )
    result = await agent.run("Check checkout-api and scale it up by one if it's under 3 replicas")
    print(result.status)   # "waiting_on_human" — paused before the mutating call
    print(result.output)

asyncio.run(main())

The run pauses with status=waiting_on_human the moment the model decides to call apply_manifest, since that tool is marked requires_approval=True. Nothing executes against the real cluster until that's explicitly approved.

Resuming a paused run

Resuming works from a completely independent process — this is the actual crash-recovery guarantee. Approving a paused run is a clean one-liner:

# Anywhere else, any time later, sharing only the same store file:
result = await agent.approve(run_id)
print(result.status)  # "completed"

Calling REST or SOAP APIs from a tool

@tool wraps any Python function, so calling an external API is no different from any other tool — there's no special Kestrion API for this:

import httpx
from kestrion.agent.decorators import tool

@tool
def get_order_status(order_id: str) -> dict:
    """Fetch order status from the orders API."""
    response = httpx.get(f"https://api.example.com/orders/{order_id}", timeout=10.0)
    response.raise_for_status()
    return response.json()

Any exception the function raises — a timeout, a 4xx/5xx, a connection error — is automatically caught and turned into a clean ToolResult.error rather than crashing the run, exactly like every other tool. What's not automatic: timeouts, retries, and secrets handling are on you to write explicitly. See examples/rest_api_tool for the patterns that matter in practice. SOAP follows the identical shape with zeep instead of httpx.

Multi-step approval chains

A tool can require approval from more than one role, not just a single yes/no:

@tool(requires_approval=["engineer", "manager"])
def deploy_to_prod() -> dict:
    """Deploys to production. Needs both an engineer and a manager to sign off."""
    ...

The run stays paused until every required role has approved — recorded via Engine.record_approval(state, "deploy_to_prod", role="engineer"), which adds a role without clobbering any already recorded (writing to scratch directly can silently destroy a partially-satisfied chain — use record_approval, not a manual dict assignment).

Time-boxed approvals

A gated tool can carry a deadline. If nobody approves in time, the run transitions to a new terminal status, EXPIRED, instead of waiting forever:

@tool(requires_approval=True, approval_timeout_seconds=3600.0)
def restart_service() -> dict:
    """Restarts a service. Must be approved within an hour."""
    ...

result = await agent.resume(run_id)            # default: status -> EXPIRED if the deadline passed
result = await agent.resume(run_id, on_expired="raise")  # or raise RunExpiredError instead

Human-in-the-Loop Input (Ask for Input)

While approval gates pause the engine for a binary yes/no decision, sometimes the agent needs a specific piece of information from a human (like an API key, 2FA code, or clarification). You can use the built-in ask_human tool (or raise InputRequired directly in your own tools).

When triggered, the engine suspends execution exactly like an approval, transitioning the run to WAITING_ON_HUMAN. You can then provide the text response back via the API to resume execution:

from kestrion.agent.tools import ask_human

# The agent invokes ask_human and the run halts.
# The user's application queries the status and prompts the human for input:
await agent.provide_input(run_id, text="My favorite color is Blue.", tool="ask_human")
# The engine records the input and resumes execution cleanly!

Vision & Multi-modal Support

Kestrion treats image inputs natively via strictly typed content blocks (TextBlock and ImageBlock). You can seamlessly pass a list of these blocks directly to the agent's run() method instead of a standard string. Kestrion automatically maps them to the correct wire format for your chosen provider (Anthropic, OpenAI, or Ollama):

import base64
from kestrion.llm.base import TextBlock, ImageBlock

with open("receipt.jpg", "rb") as f:
    base64_img = base64.b64encode(f.read()).decode("utf-8")

result = await agent.run([
    TextBlock(text="Extract the total amount from this receipt and format it as JSON."),
    ImageBlock(data=base64_img, media_type="image/jpeg")
])

Parallel tool calls

If a model requests multiple tool calls in one turn, Kestrion runs them concurrently rather than one at a time — with a safety guarantee: a batch either fully executes or cleanly pauses with nothing partially run. If any call in the batch is gated and unapproved, none of the calls in that batch run, not even the safe ones sitting alongside it.

Sub-agents

Any Agent can be wrapped as a tool another agent calls — delegation with zero new engine machinery:

specialist = Agent(provider=..., tools=[...], store=shared_store_url)
planner = Agent(
    provider=...,
    tools=[specialist.as_tool("check_inventory", "Ask the inventory specialist")],
    store=shared_store_url,  # SAME store — required for the sub-agent's run to be independently resumable
)

If the sub-agent's run pauses for approval, the parent run pauses too — the parent's scratch["_pending_approval"]["missing_roles"] will contain "sub_agent:<run_id>", naming exactly which nested run needs resuming first.

MCP client

Connect to a real MCP server and use its tools exactly like @tool functions, including approval gating:

from kestrion.mcp.client import MCPClient

async with MCPClient.stdio(command="python3", args=["my_mcp_server.py"]) as client:
    tools = await client.list_tools(requires_approval=["apply_manifest"])
    agent = Agent(provider=..., tools=tools, store="sqlite:///agent_runs.db")

MCP itself has no approval concept — requires_approval here is how you opt specific MCP tools into Kestrion's gating, by name.

MCP server

Expose a Kestrion Agent as a real MCP server so something like Claude Code can connect to it and call its full reasoning loop as a single tool:

from kestrion.mcp.server import serve_agent

agent = Agent(provider=..., tools=[...], store="sqlite:///agent.db")
mcp_server = serve_agent(agent, name="ops-agent", description="Ask the ops agent a question")
mcp_server.run(transport="stdio")

This exposes one MCP tool — ask_agent(prompt) — rather than the agent's individual raw tools. That's intentional: exposing raw tools would let a caller invoke them directly, bypassing Engine.call_tool's approval gating entirely. The full reasoning loop, including all approval gates, runs on every ask_agent call. A paused run is surfaced as a clear message (not an MCP error) so the caller knows a human approval is pending.

Pipeline / multi-agent orchestration

Run a team of agents as a DAG — some agents execute concurrently, others wait until their upstream dependencies finish:

from kestrion.scheduler import Pipeline, AgentTask, RateLimiterConfig

pipeline = Pipeline(
    tasks=[
        AgentTask("researcher_a", agent=agent_a, prompt="Research Python's history and use cases"),
        AgentTask("researcher_b", agent=agent_b, prompt="Research Rust's history and use cases"),
        AgentTask(
            "synthesizer",
            agent=synth_agent,
            prompt="Compare the two languages",
            depends_on=["researcher_a", "researcher_b"],   # waits for both
        ),
    ],
    max_workers=3,
    rate_limiter_config=RateLimiterConfig(requests_per_minute=60),  # None for Ollama
)

results = await pipeline.run()
for name, result in results.items():
    print(f"{name}: {result.status.value} — {result.run_result.output[:80]}")

researcher_a and researcher_b run concurrently in the worker pool. synthesizer only starts once both complete — dependency resolution uses asyncio.Event per task, with no polling. If one task fails, independent branches keep running (fail_fast=False default); dependents are marked SKIPPED with a clear reason. Every run is stored in the shared SQLite store and visible in kestrion dashboard.

Secure Code Sandbox

Run dynamically generated Python code in a strictly isolated, secure environment via CodeSandboxToolkit:

from kestrion.tools import CodeSandboxToolkit

# Execute code locally in an isolated temp directory or via Docker
sandbox = CodeSandboxToolkit(workspace_dir="./tmp", mode="subprocess")

agent = Agent(
    provider=...,
    tools=sandbox.get_tools() # Adds run_python_code, write_file, read_file, list_files
)

The sandbox strictly blocks path traversal (e.g. ../../etc/passwd), sanitizes environment variables to hide API keys, and automatically cleans up temporary files on sandbox.close().

Structured Outputs (JSON Mode)

Kestrion agents can return guaranteed, schema-validated Python objects instead of raw text by passing an output_schema to the Agent. We support Pydantic models natively and handle cross-provider JSON schema injection seamlessly:

from pydantic import BaseModel, Field

class Employee(BaseModel):
    name: str = Field(description="Full employee name")
    role: str = Field(description="Job title")

agent = Agent(provider=..., output_schema=Employee)

result = await agent.run("Find prime numbers up to 10")

Retrieval-Augmented Generation (RAG) Toolkit

Equip agents with native vector search capabilities using a flexible VectorStore protocol (with a zero-setup ChromaDB integration out of the box):

from kestrion.agent import Agent
from kestrion.tools import RAGToolkit
from kestrion.rag import ChromaVectorStore

store = ChromaVectorStore(persist_directory="./knowledge_base")
toolkit = RAGToolkit(store)

with open("company_policies.md") as f:
    toolkit.ingest_text(f.read(), "policies", chunk_size=1000)

agent = Agent(provider=..., tools=toolkit.get_tools())

CLI

pip install kestrion

# Scaffold a new project
kestrion init ./my-agent

# Run an agent script
kestrion run agent.py

# Generate Kubernetes manifests
kestrion deploy --target k8s --name my-agent --image registry.example.com/my-agent:latest

# Print the event timeline trace of a run
kestrion trace run_id --store kestrion_runs.db

# Output execution trace as a Mermaid flowchart diagram
kestrion trace run_id --store kestrion_runs.db --mermaid

# Time-travel debug: fork a run at a specific event sequence
kestrion fork run_id --at-seq 5

# Start an interactive terminal chat with your agent (REPL)
kestrion chat agent.py

# Launch the visual web dashboard
kestrion dashboard --port 8000

# Launch the dashboard with live-chat enabled (requires your agent script)
kestrion dashboard agent.py --port 8000

kestrion deploy generates a complete K8s manifest (Namespace, ConfigMap, Secret stub, Deployment, PersistentVolumeClaim, Service) and a Dockerfile. It never puts a real API key in the output — only a clearly-labelled placeholder. See the generated files' inline comments for what to fill in before kubectl apply.

Security & Compliance

Kestrion is designed for enterprise environments where security is a prerequisite, not an afterthought:

• Secret Management: API keys and tokens are never stored in plain text or passed in state. They are injected at runtime via the SecretProvider protocol.
• Immutable Audit Log: Every LLM request, tool execution, and state change is durably recorded in the Event Log.
• RBAC Approval Gates: Mutating actions can be gated behind role-based multi-step approval chains.
• Data Loss Prevention: The PIIRedactionMiddleware automatically scrubs sensitive PII before data is sent to external AI providers.
• Vulnerability Disclosure: See SECURITY.md for reporting policies.

For a deep dive into Kestrion's security posture, see the Security Architecture documentation.

Governance & Commercial Support

Kestrion is built to be a dependable foundation for enterprise workloads. We have structured the project to ensure longevity, transparent roadmapping, and reliable support.

• Contribution Model: See CONTRIBUTING.md for how we accept patches and govern major design changes.
• Vulnerability Disclosure: See SECURITY.md for private reporting policies.
• Public Roadmap: The living roadmap and active design discussions are available on GitHub Issues and roadmap.md.

Enterprise Support If your team is evaluating Kestrion for a production deployment, we offer dedicated design partnerships and commercial support to ensure your success.

• Enterprise Support & Inquiries: vinayak@kestrion.in

What you can build with this today

• Tool-calling agents where some actions are safe to auto-run and others need a human in the loop first — infrastructure agents, ops bots, anything touching a database or cluster.
• Multi-step approval workflows requiring sign-off from more than one role, optionally with a deadline after which the request expires.
• Agents that delegate sub-tasks to other agents, including correct approval propagation when a sub-agent's action needs sign-off.
• Agents that call tools sourced from a real MCP server, not just hand-written Python functions.
• Agents that need to survive a crash or restart mid-task. agent.resume(run_id) works from a totally different process than the one that started the run.
• Multi-turn tool use, including multiple tool calls per turn running concurrently.
• Multi-agent research pipelines and orchestration — run a team of agents as a DAG where some agents run concurrently and others wait for upstream results to arrive before starting.

Known gaps (honest, not aspirational)

• MCP is fully two-directional, both sides live-verified. kestrion.mcp.client.MCPClient connects to real MCP servers (stdio or streamable-HTTP) and is tested against a real test-fixture server, including the full approval-gating flow. kestrion.mcp.server.serve_agent() exposes a Kestrion Agent as a real MCP server, also live-verified end to end — a caller like Claude Code can connect and invoke the agent's full reasoning loop (including approval gating) as a single MCP tool. See examples/ops_demo for a worked example using both.
• Anthropic and OpenAI providers are implemented against documented API shapes but not yet smoke-tested against a live API call — no API key has been used to verify them in practice. Ollama is verified live — tests/unit/test_smoke_ollama.py and examples/ops_demo both run real agents against a real local Ollama server and pass. One real, observed limitation worth knowing: small local models can produce plain-text output describing a tool call and a plausible-sounding result without ever actually emitting a real tool-call request — Kestrion has no way to detect this, because there's genuinely no ToolCallRequest for the engine to act on; the model just wrote a paragraph claiming success. This is a model-capability limitation, not something Kestrion's approval gating or event logging can catch, since nothing was actually called. Hosted models (Claude, GPT) are far more reliable about this in practice, though unverified live as of this writing (see above).
• Multi-agent handoff is built. Agent.as_handoff_target() transfers an entire conversation to another agent, which takes over completely (distinct from sub-agents/delegation, where the original agent stays in control).
• Memory/context compaction is built. Long-running conversations are automatically summarized by the agent when history thresholds (max_history_turns or max_history_tokens) are exceeded.
• No real concurrency control across multiple agent runs. ~~Parallel tool calls within one
• Pipeline orchestration is built. The kestrion.scheduler package provides a Pipeline class with a DAG-based runner, bounded WorkerPool, and shared token-bucket RateLimiter with exponential backoff. See examples/pipeline_demo.py and the ### Pipeline / multi-agent orchestration section.
• OpenAI/Anthropic APIs. These providers are fully implemented according to the SDKs but not yet smoke-tested against a live API (only Ollama is live-verified).

Examples

• examples/pipeline_demo.py — a live demo of the Pipeline scheduler: two researcher agents run concurrently, a synthesizer agent waits for both to finish, all results saved to a shared SQLite store and viewable in the dashboard.
• examples/kubectl_agent — the original worked example, demonstrating pause-on-approval and resume-after-restart using the raw Engine/Node primitives directly (useful for understanding what Agent builds on top of).
• examples/rest_api_tool — patterns for calling REST/SOAP APIs from a tool: explicit timeouts, gating a mutating call, reading secrets from the environment, and writing your own retry loop.
• examples/ops_demo — an integration demo exercising parallel tool calls, sub-agent delegation, a multi-role approval chain with a timeout, and multi-agent handoff together against a real local Ollama model. Run live, this also surfaced a real limitation: a small local model can describe a fabricated tool call and result in plain text without ever emitting a real tool-call request — see the Known Gaps note above.
• tests/unit/test_smoke_ollama.py — a live, real smoke test against a local Ollama server. Skips automatically if Ollama isn't running.
• tests/unit/test_mcp_client.py — a live test against a real MCP server (tests/fixtures/mock_mcp_server.py), including the approval-gating integration.
• examples/mcp_server.py — a worked example exposing a Kestrion agent as an MCP server over stdio.
• tests/unit/test_mcp_server.py — a live test of the MCP server side (serve_agent()), verifying a paused run is correctly surfaced through the MCP protocol rather than misreported as an error.
• tests/unit/test_cli.py — CLI integration tests for kestrion init, kestrion run, and kestrion deploy --target k8s, including the security check that generated manifests never contain real-looking API keys.

Documentation

• Getting Started
• Architecture
• Concepts: Event Sourcing · Checkpointing · Approval Gates · Sub-Agents vs. Handoff

Development

git clone https://github.com/VinayakDubey07/kestrion.git
cd kestrion
python3 -m venv .venv
source .venv/bin/activate
pip install -e ".[dev]"
pytest tests/ -v
ruff check src/ tests/

Roadmap

Next up: Further hardening, OpenTelemetry exporters, and Human-in-the-Loop "Ask for Input". The scheduler and Postgres store are now built — see src/kestrion/store/postgres_store.py and src/kestrion/scheduler/. See roadmap.md for the detailed, dated 3-month plan.

License

Apache 2.0