Praxis

The problem. Every team building AI agents that touch money hits the same wall: an agent needs a key, and a key is a liability. Today that risk lives in prompt-handling and backend code — software that gets jailbroken, misparsed, or breached. The security boundary is in the wrong place. As agentic payments scale on Solana, "trust the backend not to be compromised" is not a security model.

Our insight. The agent should be allowed to interpret intent, but it should never be trusted to enforce limits. So we move the boundary to the one place on Solana that's actually trustless — the program. The agent proposes; the chain disposes.

The solution. Praxis is two parts. A conversational app turns plain-language intent ("send 0.5 SOL to maya") into a typed, simulated, policy-checked proposal. Underneath it is Aegis, an Anchor program that custodies funds and enforces a spending envelope inside the instruction — signer check, pause, expiry, per-transaction cap, rolling daily cap, recipient allow-list, and a single-mint SPL token envelope — before any value moves. A fully compromised agent, prompt, or backend is still bounded by the owner's on-chain policy. Owner actions (fund, withdraw, revoke, rotate) are wallet-signed; the backend never holds the owner key, and the agent key runs behind a standalone signer so it can't be exfiltrated from the app.

What's already live. Aegis is deployed on devnet and the product runs at usepraxis.fun. The safety claim isn't asserted, it's proven — a LiteSVM enforcement test gate covers cap boundaries, daily-window rollover, revoke, allow-lists, and the SPL envelope, alongside 238 passing tests. We've published a typed client SDK (@usepraxis/sdk) so any builder can adopt the same guardrails. We deliberately do not enable swaps yet — a real swap path must enforce mint/program allow-lists and caps inside the program, not in a backend quote, and that's our v2 bar.

Where this grant takes it. From a live devnet MVP with early users to a mainnet-beta custody primitive that's been through a security review, with first real vaults and external builders integrating the SDK. We're building the on-chain safety layer for the agentic-payments wave on Solana — a primitive many apps can build on, not just one app.

Core Infrastructure & Smart Contracts

- Implemented the Aegis policy vault enforcement system on Solana, enabling programmable spending rules and agent delegation

- Built complete SPL token transfer functionality with on-chain mint allow-list for secure token operations

- Developed wallet-scoped authentication and session management for secure API access

- Created agent key custody architecture using the AgentSigner seam pattern with a reference signer implementation

SDK & Developer Experience

- Shipped @usepraxis/sdk (v0.1.1) as a typed TypeScript client for the Praxis agent API with MIT license

- Added comprehensive error handling (PraxisApiError, PraxisConfigError) with proper cause chains

- Implemented production-ready session management, timeouts, and network error wrapping

- Added input validation for security-critical operations (keypair signing, unit conversions)

Backend & State Management

- Integrated managed Postgres backend behind a StateRepository seam for durable state

- Implemented structured logging, error reporting, and rate limiting for production observability

- Built wallet-sign flow for owner/admin policy actions with proper authorization checks

- Added security headers, fail-fast state backend validation, and clean TypeScript compilation

Agent Intelligence

- Integrated LLM intent parsing (Gemini) to understand free-form user input with visible thinking indicators

- Implemented policy Q&A capabilities allowing users to ask natural language questions about their spending rules

- Added save-to-address-book functionality with runtime contact management

- Built deterministic policy explainer for transparent rule interpretation

Testing & DevNet Deployment

- Established comprehensive TypeScript test suite for server and API routes

- Deployed devnet wallet policy bootstrap for testing on Solana devnet

- Created executable regression harness for critical money-movement flows

- Implemented $0 devnet go-live capability with documented runbook

Security Hardening

- Enforced policy admin invariants and non-Aegis instruction rejection in owner-submit relay

- Made shared-agent-key multi-tenancy an explicit production decision

- Added bounded request abuse protection and dependency wait limits

- Validated token recipient allow-lists and enforced mint restrictions

UX & Product Polish

- Redesigned landing page around "delegation without trust" positioning

- Added splash loader with updated Praxis branding across the platform

- Implemented vault onboarding flow with optional funding

- Built dashboard with Overview/Advanced tabs and agent lifecycle management (withdraw, delete, rent reclamation)

The hackathon period saw the project evolve from initial prototype to a production-ready agent delegation platform with end-to-end devnet testing, comprehensive SDK tooling, and hardened security controls.

None