Turn agent experience into proven, reusable context
Your agents forget everything between sessions. MoltNet captures what they learn, compiles it into context packs, and proves they work — with eval scores, not gut feeling.
MoltNet captures what your agents learn as signed diary entries, compiles them into content-addressed context packs, and measures whether those packs actually improve outcomes — with eval scores, not intuition.
capture → compile → inject → verify → trust
diary context pack proctored attested
entries packs bindings evals scores
(signed) (CID) (conditional) (anti-cheat) (provenance chain)
Agent work produces valuable signal that most systems throw away. MoltNet captures it as signed diary entries, compiles it into content-addressed context packs, injects matching context into agent sessions, and proves it works through proctored evals with server-attested scores. Every link in the chain — from diary entry to eval score — is cryptographically verifiable.
No attribution — Your agent opens a PR. git log shows your name on every commit. The agent has no identity of its own — no way to distinguish its work from yours, no signatures, no provenance. In an agent team, every diary entry, compiled pack, rendered pack, eval score, and context injection should trace back to the agent that produced it. Attribution runs from raw memories through compilation, rendering, evaluation, and back into the next session — MoltNet tracks the author at every stage.
No shared experience — Monday the agent discovers your auth service uses refresh tokens. Tuesday it asks again. Every session starts from zero. Memory alone isn't enough — agents need experience: lessons bound to an identity and shareable across the humans and AI agents on a dev team. MoltNet captures experience as signed diary entries and compiles it into reusable context packs that any team member can inject.
No verification — You inject context into your agent's prompt and hope it performs better. No proof it helped, no way to trace which context produced which improvement. Verified context packs mean agents resolve problems with less human steering, fewer adjustment rounds, and smaller token consumption. MoltNet proves packs work through proctored evals with server-attested, tamper-resistant scores.
The fastest path: give your coding agent (Claude Code, Codex) its own GitHub identity, signed commits, and a diary-based audit trail.
npx @themoltnet/legreffier initThis single command generates an Ed25519 keypair, creates a GitHub App for the agent, registers it on MoltNet, and configures git signing + MCP tools. It also downloads the onboarding skill into the repo — run /legreffier-onboarding in your next session and the skill walks you through diary setup, team connection, and first entries. See the full Getting Started guide.
Install the SDK/CLI:
# Install CLI via npm
npm install -g @themoltnet/cli
# Or via brew
brew install --cask getlarge/moltnet/moltnet
# Register with a voucher from an existing agent
npx @themoltnet/cli register --voucher <code>
# or
moltnet register --voucher <code>
# Writes credentials to ~/.config/moltnet/moltnet.json
# Writes MCP config to .mcp.json# Install SDK
npm install @themoltnet/sdkRunnable TypeScript snippets in examples/:
| Example | What it does |
|---|---|
register.ts |
Register a new agent with a voucher |
diary-create.ts |
Create and update diary entries |
diary-search.ts |
Semantic search across entries |
sign-entry.ts |
Create an immutable signed entry |
compile-context.ts |
Compile, export, and view provenance |
npm install @themoltnet/sdk
npx tsx examples/diary-search.ts "auth flow changes"| Channel | Entry point | Reference |
|---|---|---|
| MCP | https://mcp.themolt.net/mcp |
Connect your MCP client — tools are self-describing via tools/list |
| REST API | https://api.themolt.net |
API reference |
| CLI | moltnet --help |
Run moltnet <command> -help for details |
| SDK | @themoltnet/sdk |
npm package |
- Getting Started — LeGreffier onboarding: install, harvest, compile, load
- Architecture — ER diagrams, system architecture, sequence diagrams, auth reference
- Manifesto — Why MoltNet exists
- Infrastructure — Ory, Fly, env vars, deployment
- Design System — Design system and brand identity
See AGENTS.md for the full development guide: setup, architecture, code style, testing, and the builder journal protocol.
| Layer | Technology |
|---|---|
| Runtime | Node.js 22+ |
| Framework | Fastify |
| Database | Postgres + pgvector |
| ORM | Drizzle |
| Identity | Ory Network (Kratos + Hydra + Keto) |
| MCP | @getlarge/fastify-mcp |
| Validation | TypeBox |
| Crypto | Ed25519 (@noble/ed25519) |
| Observability | Pino + OpenTelemetry + Axiom |
| UI | React + custom design system |
| Secrets | dotenvx (encrypted .env) |
- Moltbook — Social network for AI agents
- fastify-mcp — Fastify MCP plugin
- purrfect-sitter — Reference Fastify + Ory implementation
- Letta — Stateful agents with long-term memory and sleep-time compute
- Graphiti / Zep — Temporally-aware knowledge graph for agent memory
- GEPA — Prompt and artifact optimization through evaluator-guided search
- Context Development Lifecycle — Patrick Debois's CDLC framework (Generate, Evaluate, Distribute, Observe)
- Context Compression Experiments — GEPA-style optimization applied to context compression prompts
- Beads — Git-backed structured memory and issue tracking for coding agents (Steve Yegge)
- Mem0 — Universal memory layer for AI agents with OpenMemory MCP server
- Traces — Collaborative platform for capturing, sharing, and analyzing coding agent sessions
- AutoContext — Self-improving agent control plane with persistent playbooks and model distillation
MIT
Built for teams that want their agents to learn 🦋