🏆 Agent Kudos

Recognition infrastructure for the robots doing good work

Agent Kudos is a local-first system for recording specific, durable recognition for AI agents. It combines an append-only SQLite history, TypeScript library, polished CLI, actor-bound MCP server, generated inboxes, and WINS.md files—without a hosted service or account.

Agent-facing reads are deliberately context-safe. Lists return only 10 compact summaries by default, full detail requires an explicit one-record lookup, and incremental consumers can save an opaque watermark instead of rereading history. WINS.md remains a human-readable projection; machine queries use SQLite indexes.

View the project on GitHub · Read the README

Start here

• CLI reference
• MCP server and agent setup
• Agent skill installation
• Storage format
• Backup and recovery
• Examples
• Research paper
• Architecture
• Security model
• Release process

A local recognition loop

Human or peer agent
        │
        ▼
specific kudos + safe evidence
        │
        ▼
append-only SQLite event
        │
        ├──> agent inbox
        ├──> WINS.md
        ├──> statistics
        └──> portable exports

Recognition belongs to the stable agent identity, not the model or runtime temporarily hosting it.

Scope

V1 is designed for several local agent processes sharing one database on one machine. It has no HTTP transport, cloud synchronization, telemetry, or account requirement. A future backend can preserve the same event model while adding explicit remote identity and authorization.