Alpha 6 made each agent yours to shape. Alpha 7 widens the world they live in. You can now bring any Claude or Gemini API key directly (no CLI, no subscription), install MCP servers from a catalog with OAuth handled for you, send and receive on WhatsApp and Signal, and watch agents turn their own memories into a queryable knowledge graph. Permissions, integrations, and long sessions all got a steadier hand.
What's new
🔑 Native API providers
Use Anthropic and Google directly with your own API keys, side by side with the CLIs you already had. Two new CLI providers join too: Codex and Gemini CLI. That brings the working list to: Anthropic API, Claude Code CLI, Google Gemini API, Gemini CLI, Codex CLI, OpenAI, DeepSeek, GLM, Kimi, Mistral, Ollama, Qwen, LM Studio, Docker Model Runner. Pick what fits the job; agents stream and call tools the same way regardless.
🧩 MCP servers, properly managed
- Install wizard: pick an MCP server, fill the form, run.
- Per-agent bindings: enable a server for the agents that need it, leave others untouched.
- Secrets, isolated per binding: API keys stay scoped, never leaked into prompts.
- A supervisor watches the subprocesses so idle servers get reaped and busy ones stay healthy.
- Bundled Node.js and uv runtimes so MCP servers that need them just work, no global installs.
🧠 Knowledge-graph memory
Agents now extract structured subject → predicate → object facts from their own memories in the background. You get the running prose memory you already had and a graph you can query. Each agent picks its own extraction model and prompt, or uses your global default. Long sessions also gain a rotation lifecycle: when context gets tight, important facts are rescued into memory before summarization kicks in, instead of being lost on the cut.
📲 More ways to reach your agents
- WhatsApp: link your phone with a QR code; agents can send and receive messages.
- Signal: dedicated linking flow with a managed sidecar.
- Webhooks: send and receive structured payloads with verification and clear error toasts.
- Email: rebuilt on a modern IMAP stack with TLS/STARTTLS, real-time IDLE watching, and saner failure handling.
- Discord channel discovery via your own bot credentials.
- Slack mode refinements for cleaner messaging.
🎚 Per-agent subscriptions
Each agent now decides, per integration:
- whether to wake on a new message or stay silent until mentioned,
- whether to show an outbound notice when it's the one starting the conversation,
- which channels it speaks for at all.
No more "every agent answers everything" by default.
🛡 Permissions that remember
- Normalized grants describe what an agent can do (capabilities and selectors), not just which tool calls slipped through.
- Persistent settings: your decisions stick across sessions, with auto-approval available per agent.
- Stale segments get marked when grants or tools change, so reviews stay honest.
- Unattended handler for non-interactive providers, so background runs don't hang on a prompt nobody sees.
🪟 UI polish
- New PromptField input across panels (subscriptions, schedules, skills, instructions, memory, tasks) replaces the older textarea everywhere.
- Tooltip rewritten to position correctly even inside scrolling panes and modals.
- Toggle switches refined for accessibility and contrast.
- Provider settings pages get proper logos and live-update when you enable or disable a provider elsewhere.
- Setup guide collapsed by default; placeholder and help text clarified across settings.
📚 New starter templates
Stats & Growth Reporter, Customer Support Triagist, Technical Writer, Vendor & Resource Scout, Weekly Reflector, Product Analyst. Plus a new append-to-system-prompt field for templates so they ship with their personality without forking.
Under the hood
- Agent creation gated on having an account: if a CLI binary is detected but no Apprentice account is configured for it, the empty state now points you to settings instead of half-creating an unusable agent.
- Container lifecycle smoothed: agent mounts pre-register before first start, status polling is throttled, and host timezone is detected and propagated into shared containers.
- Codex and Gemini CLI can run one-shot commands in-container with proper auth handling.
- readMachineID hardened on Windows to stop the brief console flash.
Known limitations
Still alpha. Expect rough edges, breaking changes before v1.0, and the occasional bug. The MCP modification is one of the latest and complex additions. Knowledge-graph extraction runs as a background queue, so the graph for older memories fills in over a few minutes after upgrade.
Report issues via GitHub Issues.