Skip to content
Bestie AgentBestie Agent

Roadmap and priorities

Bestie’s roadmap is sequenced to avoid adding dangerous autonomy before the local runtime, memory controls, Doctor, channel behavior, and permission layer can carry it.

Dangerous mistakes to avoid

Section titled “Dangerous mistakes to avoid”
  1. Do not build Zep before local SQLite.
  2. Do not broaden MCP/ACP before Doctor, permissions, and local tool logging are boringly reliable.
  3. Do not let playful rude tone become abusive.
  4. Do not market Bestie as conscious, therapy, romance, or perfect memory.
  5. Do not make onboarding v1 too long.
  6. Do not turn controlled autonomy into silent risky automation.

Available now

Section titled “Available now”

Implemented in the local MVP foundation:

  • terminal chat
  • editable character prompt and installed skills
  • onboarding/config wizard
  • config v2 provider profiles and model refs
  • LLM setup paths for OpenAI/ChatGPT, Anthropic Claude, Groq, OpenRouter, custom compatible endpoints, Ollama, and native Gemini API-key mode
  • redacted operational logs
  • local SQLite memory with inspect, search, pending approval, hygiene, governance, pause/resume, export, and cleanup flows
  • local knowledge graph with entities, relations, pending graph writes, review suggestions, and hygiene analysis
  • Doctor diagnostics and safe local fixes
  • channel registry and bestie channels command group
  • Telegram local polling with owner-only access, slash commands, attachments, progress messages, memory controls, and approvals
  • Zalo text-only local polling with owner-only access, slash commands, diagnostics, memory controls, and approvals
  • cron schedule CRUD and scheduler runtime
  • daemon management for telegram, zalo, cron, or all
  • cross-platform user service integration through Linux systemd, macOS launchd, and Windows Startup
  • permission-gated internal tools for local reads, writes, patches, exec, process listing, web reads, git, memory, knowledge, and cron management
  • bounded internal subagent spawning for focused helper investigations
  • SDK-backed MCP add/list/show/test/tools/classify/login/call flows with OAuth setup and classified read calls
  • local web console through bestie ui
  • npm update checks and bestie update --apply
  • character regression evals and smoke checks

Now: Action Inbox Core

Section titled “Now: Action Inbox Core”

The next product milestone should turn the existing runtime into a convenience layer.

Build the smallest Action Inbox Core:

  • define the ActionItem data model
  • collect candidate items from chat, cron reports, memory, knowledge graph, and tool results
  • rank items by urgency, value, risk, deadline, and blocked-workflow impact
  • render a short Daily Command Brief for Telegram and terminal
  • support owner commands: approve, delay, ignore, edit first, remember, and never again
  • learn repeated decisions into memory or rules when safe
  • keep public, external, destructive, and money actions behind explicit approval

The goal is not more dashboard. The goal is fewer decisions for the owner.

Next: Reliability and onboarding polish

Section titled “Next: Reliability and onboarding polish”

Good follow-up work:

  • tighten first-run onboarding when it improves completion
  • improve Doctor diagnostics for real setup failures
  • reduce Telegram/Zalo edge-case failures from daily usage
  • make cron reports more actionable and less noisy
  • improve local web console flows only where they reuse runtime services
  • improve memory/knowledge hygiene so the agent does not become a context landfill

Later

Section titled “Later”
  • optional Zep memory
  • backup, restore, migration, and portability
  • richer local web console and Doctor UI
  • broader tool surfaces after permission defaults have real-channel mileage
  • richer channel adapters after owner-only polling flows are stable
  • productized UI for Action Inbox and approvals

Future

Section titled “Future”
  • broader MCP execution categories after classification, logging, and approval UX are mature
  • named ACP/multi-agent orchestration beyond bounded helper subagents
  • plugin system
  • persona templates and marketplace
  • avatar, voice, and body layer
  • hosted/SaaS mode

Hard rules

Section titled “Hard rules”
  • SQLite before Zep.
  • Doctor and permissions before broad external tools.
  • CLI/runtime services before UI duplication.
  • Installer after onboarding works.
  • Safe preparation can be automatic; risky action requires approval.
  • Public claims must avoid consciousness, therapy replacement, romance, and perfect memory.