roadmap

Direction, comparison & changelog.

Where Manef is heading, how it sits against the field, and what has shipped. If you want to help build it, pick a slice from the roadmap and open a PR.

where it's going

Not another agent harness — a foundation.

OpenClaw owns channels, 9Router owns routing, Hermes owns the learning loop. Manef isn't racing them on feature count — it's the layer underneath: multi-tenant, encrypted per-user, and composable, where every capability ships as a droppable rr slice you add to any app with one command. The roadmap fuses the best of all three, then builds what none of them ship.

Shipped — the foundation

  • Multi-tenant BYOK gateway · 22 providers
  • Per-user AES-256-GCM keys, host-gated
  • Agent loop + saved agents + full traces
  • Memory + Obsidian-style graph
  • MCP server (OAuth 2.1 PKCE)
  • MCP client — connect external servers
  • /v1 OpenAI + Anthropic gateway
  • 4 inbound chat channels
  • Scheduled agents · usage · caps · audit
  • Multi-key failover pool · cooldown/backoff
  • 13 droppable rr vertical slices

Next — adopt from the field

  • OAuth-auth external MCP servers
  • Vision + embeddings → RAG
  • Streaming /v1 + tool passthrough
  • More channels + a cost / quota dashboard
  • Workspace-shared key pool · more OAuth logins
  • CLI / TUI surface

The vision — what none of them ship

  • One policy engine — approval · sandbox · budget · scope
  • One observability timeline — channel → memory → model → cost
  • Portable memory + skill contract across runtimes
  • Task-aware cost governance
  • First-class eval loop — outcome · feedback · regression

Every item is a self-contained vertical slice with a typed contract — build it once here, drop it into any app. Read the roadmap ↗ · pick a slice & contribute ↗

changelog

What shipped, when.

2026-07-10admin + landing + isolation
  • Admin: analytics charts, per-user activity log + exact signup time, JSON preset seeding
  • Landing: roadmap + changelog moved into a shared header; de-duped hero CTA
  • Security: external MCP tool surface now workspace-scoped (was per-user)
  • Roadmap corrected: MCP client + provider-pool failover marked shipped
  • BYOK: register >1 key per provider — the failover pool is now populatable
2026-07-08landing + hardening
  • Memory-graph slice + Obsidian graph; icon-rail app shell + mobile dock
  • CRUD completeness: updateCombo, mcp updateServer, schedule edit, channel cascade-delete
  • Security: invite privilege-escalation closed; channels setModel wired
  • Landing rebuilt: live graph demo, this roadmap + comparison + changelog
2026-07-07audit fix pass 1
  • Spend-cap fails closed on read truncation (was silent under-enforcement)
  • workspace transferOwnership; 402/quota failover; combos round_robin; audit hooks
2026-07-05verification audit
  • 5-agent per-phase audit vs code; spend-cap runAgent bypass fixed
  • Best-practice + CRUD scorecard: 0 HIGH issues, avg ~87/100
2026-07-04v0.1 shipped
  • 5 phases live: workspaces · BYOK /v1 gateway · memory · 4 channels · scheduled agents · usage/caps/audit
  • 13 rr vertical slices; per-user AES-256-GCM keys; MCP server (OAuth 2.1)
how it compares

Distilled from the best — multi-tenant by design.

Manef borrows the strongest ideas from OpenClaw (channels / gateway), Hermes (memory + agent loop) and 9Router (routing + token savers) — then adds what none of them have: real per-user tenancy and encrypted-at-rest keys.

FeatureOpenClawHermes9RouterManef
Per-user multi-tenancy
Keys encrypted at restplaintextplaintextplaintextAES-256-GCM
BYOK providers~60289422
Agent loop + saved agents
Memory + interactive graph
MCP server (OAuth 2.1)
OpenAI / Anthropic /v1 gateway
Chat channels (Telegram / Slack …)

Only Manef ships true per-user multi-tenancy + AES-256-GCM keys at rest — the others are single-user with plaintext credentials. Full comparison ↗

go deeper

The full docs.