Flowly 3.3.0 — Agent connections and reliable tool activity
NewManage MCP and Gmail connections per profile, with browser authorization, identity-bound setup and authenticated remote management.
NewEdit agent memory and profile information through the shared app interface; steer active conversations without stopping running tools.
ImprovementsMCP connections recover from interruptions, preserve rich results and keep tool access scoped to the conversation and agent.
FixesTool progress and delegated results stay consistent across clients; group output preserves Unicode, Gmail pagination and message reading are corrected, and Android profile sessions are listed correctly.
Flowly 3.2.0 — Bots
NewBots: run more than one agent on the same machine. Each keeps its own memory, settings, skills and credentials, and cannot read another's.
NewBot groups: put two to six bots in one conversation. Choose who answers every message and who waits to be @mentioned, and whether they answer all at once or one after another.
NewA group keeps its whole history and pages older messages in as you scroll, counts what each member spends, and comes back idle after a restart instead of showing a turn that stopped being written.
NewA bot can hand work to another bot, addressed by the name you gave it rather than the identifier it has on disk.
NewBots travel: export one as a sealed archive and import it on another machine, colour and mark included.
ImprovementsA running gateway says what it is on /health, so an app never has to guess whether the agent on this machine is one it started.
FixesA cache hit is billed at the cache price, not the input price. A group no longer sweeps away a file one of its messages still points at. A failed request says something you can act on instead of naming a file on your disk.
NewFlowlets: your agent now builds you live mini-apps on request — trackers, tables, rich charts, forms, timers, even photo capture with vision — as real interactive screens wired to live data.
NewFlowlet reminders & watches: declarative triggers that fire a notification when your data changes — no model call, no scheduled prompt.
NewPlan mode: for bigger tasks the agent proposes a plan first, you approve, then it acts.
NewLive permission control in the TUI: Tab / Shift+Tab cycles Ask, Auto, and YOLO with a colored status-bar badge.
NewExec and Codex approval policies are now editable live from any connected app over the shared RPC.
NewEmbedded browser provider in the gateway — the engine behind Desktop's new built-in browser.
NewFlowlet templates: ready-made mini-apps — sleep and mood tracking included — as a one-tap start.
NewStanding goals: hand your agent a goal and it keeps working toward it on its own — waiting on a judge or a barrier, surviving a restart, and telling you when it's done.
NewGoals where you work: a composer strip shows what's being pursued, slash commands set it, pause it and stop it, and autonomous turns stream into the chat as they happen.
NewWatch a scheduled job while it runs: each run streams its actions live, then records what it did and every file it produced.
NewA full tool catalog behind tool_search: a turn loads only the tools and playbooks it actually needs, with optional on-device routing that requires no API key.
Improvementsthe CLI starts a fresh session by default (pick up an old one with --resume); chat artifacts appear right under the TUI input; /usage is a true inline panel.
Improvementsweb search now works out of the box for signed-in accounts, with your own key still taking priority.
Improvementsplan mode endures long sessions: plans survive context compaction and gateway restarts; flowly doctor was rebuilt on a read-only core with transactional repairs.
Improvementslong sessions hold together: the gateway reads a model's real context window, learns a provider's true input ceiling from its own rejections, and carries task intent and recent events across compaction.
Improvementsevery surface serializes an approval the same way, so a request raised in one place can be answered from another.
Fixesthe loopback WebSocket is hardened against web-origin and DNS-rebinding attacks; deleted sessions release their pinned working directory; xAI errors now show the real reason from xAI.
Fixesconnections reveal their secrets safely and accept restricted ElevenLabs keys; a tool whose server changed its schema refreshes instead of failing.
NewFlowly's agent core is open source (Apache 2.0): the gateway, 40+ tools, 135 built-in skills, and every channel adapter — the same code that ships inside Flowly Desktop, with no separate "lite" build. Install with uv tool install flowly-ai or the one-line script.
NewBring your own key: first-class adapters for OpenRouter, Anthropic, OpenAI, Google Gemini, Groq, xAI/Grok, and Zhipu/GLM, plus any OpenAI-compatible local model (Ollama, LM Studio, vLLM). When nothing is pinned, Flowly cascades through whatever you've configured so there's always a working model. Switch live with /provider and /model.
NewNative Anthropic provider: direct Messages API with prompt caching and automatic key rotation, not just an OpenAI-compatible shim.
NewxAI subscription sign-in: use your SuperGrok / X Premium+ subscription instead of an API key.
NewOne gateway, every channel: terminal TUI, Telegram, Discord, Slack, Microsoft Teams, WhatsApp, iMessage, email, and voice — one conversation memory shared across all of them.
NewMCP both ways: connect any MCP server (flowly mcp install) and run Flowly itself as an MCP server (flowly mcp serve) for Claude Desktop, Cursor, or Claude Code.
NewA closed learning loop: every fact becomes a governed memory with a calibrated trust score; a background pass merges duplicates and retires stale notes; structured facts land in a knowledge graph. Opt in and Flowly writes and refines its own skills — every change snapshotted and reversible.
NewDelegates and parallelizes: spawn isolated sub-agents, run a cross-channel task board sequentially or in parallel, and hand heavy coding to a local Codex session (opt-in).
NewRun it anywhere: self-host on a laptop, a $5 VPS, or a Mac mini as a background service that survives reboots (launchd / systemd / Task Scheduler). The whole process runs inside the OS sandbox (sandbox-exec on macOS, bubblewrap on Linux).
NewInline reply media: generated images and screenshots now render right in the chat across terminal, desktop, and mobile — a small thumbnail you tap for full resolution.
ImprovementsFull documentation at useflowlyapp.com/docs, sourced directly from the public repo: every channel, provider, and tool, plus self-hosting and the open-source-vs-Desktop/Cloud split.
InternalPublished to PyPI as flowly-ai; engineering deep-dives (memory governance, skill self-improvement, internal gateway RPC) live in the repo.