Model providers
Configure an OpenAI-compatible hosted or self-hosted model provider. RetifyOS calls it through the agent runtime and passes tool access through the application’s registered capability set.
OPEN-SOURCE AGENT INFRASTRUCTURE
Models bring the reasoning. RetifyOS brings the runtime around it: registered capabilities, persistent run state and a defined path from an agent’s intent to controlled execution.
An agent needs more than a prompt and an API key. It needs systems that can turn a request into a bounded tool call, keep track of the work and make high-impact actions follow explicit rules. RetifyOS brings those operating foundations together in an open-source project developers can inspect, run and extend.
An agent needs more than a prompt and an API key. It needs systems that can turn a request into a bounded tool call, keep track of the work and make high-impact actions follow explicit rules. RetifyOS brings those operating foundations together in an open-source project developers can inspect, run and extend.
01 WHY AN OPERATING LAYER
A model can interpret a request and propose what to do next. The systems around that reasoning have different responsibilities: deciding which tools are available, preserving run state, tracking what happened and applying operational rules consistently.
When every agent rebuilds those foundations around its own prompt and API wrapper, the result is fragmented and difficult to inspect. RetifyOS gives developers a shared runtime beneath the model. The reasoning provider can change; tool access, run state and execution controls remain explicit parts of the system.
Read the architecture02 THE RUNTIME
RetifyOS connects the model, registered tools and execution path without giving the model authority over the systems that enforce it.
Configure an OpenAI-compatible hosted or self-hosted model provider. RetifyOS calls it through the agent runtime and passes tool access through the application’s registered capability set.
Tools declare their schemas, effects, timeouts and required capabilities. The runtime exposes registered tools through a dispatcher that checks the caller’s grants; the same registry can serve agent calls and MCP clients.
The application stores run state and events in SQLite alongside sessions, then exposes run events through resumable API streams. Operators and clients can inspect and reconnect to the work—not just receive a final model response.
The repository includes an authenticated API, local or remote CLI, MCP stdio access, an optional Telegram adapter and an operator console. These are interfaces to configure and run from the repository, not a claim that RetifyOS is a hosted public service.
03 MORE IN THE CODEBASE
The repository also contains a durable memory store, filesystem skill manager, context-compilation utilities and a SQLite-backed job queue. These modules cover revision-checked memory, skill discovery and loading, context pruning with sensitive-value redaction, and scheduled or event-triggered jobs with leases, retries and dead-letter handling.
Integration status: These are concrete project modules available for integration. The default createApplication composition does not instantiate or connect them.
INTERACTIVE WALKTHROUGH / LOCAL DEMO
Follow a proposed action through a simulated check and an explicit operator decision. This walkthrough makes the control boundary visible, one step at a time.
Offline walkthrough. It uses local illustrative data; it does not connect to an API, wallet, network or signer and cannot execute or approve a transaction.
A model or another upstream process may suggest an action. Here, the proposal stays separate from execution.
The demo pauses to show where a deterministic check could inspect a proposal. No external system is contacted.
The final card is a simulated approval checkpoint. Pressing the button does not authorize or submit anything outside this page.
The walkthrough stays on this page; no external service or transaction is involved.
04 THE EXECUTION BOUNDARY
In the documented Solana/Jupiter swap lifecycle, a typed request becomes an exact transaction, checked against policy and simulated before operator approval.
Independent signing boundary: approval creates a short-lived authorization envelope. A separate signer verifies it before signing; the model/API process does not receive signer key material.
Trading docs ↗05 FOR DEVELOPERS
RetifyOS is an open-source project with a local verification flow. Clone the repository, install the dependencies and run the project checks before configuring your own model or network connections.
Requirements: Node.js 22 or newer. The verification suite runs offline against fakes and local SQLite; it checks repository logic, not live-chain behavior or an independent security audit.
Read the Operations Guide$ git clone https://github.com/RetifyOS/RetifyOS.git
$ cd RetifyOS
$ npm ci
$ npm run verify
06 CURRENT IMPLEMENTATION & DIRECTION
The repository composes a configured model provider, capability-aware tool registry, durable run/session stores, API, CLI, MCP and Telegram interfaces, plus an operator console.
The transaction lifecycle includes quoting, simulation, policy and risk checks, approval, separate signing, broadcasting and reconciliation.
Robinhood Chain remains in development; it is not a live integration.
Live Solana execution is disabled by default and requires explicit operator configuration and approval.
OPEN-SOURCE AGENT INFRASTRUCTURE
Start with the runtime, verify it locally, then extend the system around the agent you want to build.
Read the Architecture