You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Add several connectors per provider. You can now configure up to 10 named connector instances — e.g. two OpenAI-compatible endpoints (a local model and a cloud one), or separate Gemini keys for work and personal — each with its own API key, model slots, and tuning.
Pin a conversation to a connector. The provider/model switcher pins the current tab to a connector instance; other tabs and new sessions keep using your global default. Per-connector API keys are stored individually in the keychain.
Per-instance feature flags. Watch-word recovery, thinking, and tool-use settings now follow the connector a session actually resolves to, not the global settings.
Skill management
In-app skill editor. Create and edit skills (instructions, metadata, scripts) directly from Hobbes — no more hand-editing files.
/command invocation with autocomplete. Type / at the start of a line to invoke a skill by name, with a searchable autocomplete popover. You can put context on the lines above the command; mentioning a skill mid-sentence stays plain text.
Live reload. A file watcher picks up external changes to your skills folder immediately.
SQLite session storage
Chat history moved from sessions.json to a database. Sessions load lazily (open tabs and the active session hydrate at startup; everything else loads on demand), saves write only what changed, and nothing ever expires. Your existing history is imported automatically on first launch — the old JSON files are left in place as a backup.
Faster history search. Search runs in the database across all sessions — including ones never opened this run — matching names, summaries, and message text.
Composio
No-auth toolkits connect cleanly (e.g. tools that need no account link).
Remove toolkits and recover wedged servers from MCP → Status.
Profile isolation. Clients and on-demand tools are scoped to the active profile, so tools no longer bleed across tabs using different profiles.
🛡️ Reliability & Security
Session-store hardening: saves are atomic (a crash can't persist a half-written state), overlapping background saves can't revert newer data, deleted conversations stay deleted, and a corrupt legacy sessions.json is preserved for recovery instead of being treated as empty history.
Keychain fixes: key-discovery indexes are always readable before authentication, and keys migrated from older versions honor your chosen storage mode — no more surprise Touch ID prompts at startup or being bounced back into onboarding despite a valid key.
Model pinning safety: deleting a connector no longer sends its pinned model name to a different provider.
UI stability: fixed a potential crash when autocompleting skills in text containing emoji/CJK, streaming stays live when returning to a tab mid-stream, and the session cost icon refreshes on stream lifecycle.
Dependency audit refresh: dependencies bumped and the RUSTSEC ignore list pruned from 10 entries to 3, each with a documented justification.
Smart tool selection no longer depends on Gemini. When you connect a Composio toolkit with more tools than the model limit (128), Hobbes asks an LLM to pick the most useful subset. That step was previously hardcoded to Gemini, so connecting a large toolkit failed outright if you didn't have a Gemini key. It now uses your active provider (Gemini, Claude, or any OpenAI-compatible endpoint), falling back to any other configured provider.
All providers implement tool selection — Claude (Messages API), OpenAI Chat Completions, and OpenAI Responses join Gemini, each using its configured summary model.
Edit a toolkit's tools directly from Hobbes
New tool editor in MCP → Status. Composio removed the easy in-dashboard way to enable/disable individual tools for a connection, so Hobbes now has its own. Each connected toolkit gets an edit button that opens a searchable checklist of every tool the toolkit offers.
Search, bulk toggle, and live counts. Filter a long tool list by name or description, check/uncheck everything currently shown, and watch an enabled-count badge that warns when you go above the optimal (50) or provider (128) limits.
Saves apply immediately. Your selection is written as the connection's tool whitelist and the tool set is hot-reloaded — caches are busted (toolkit counts, on-demand discovery, and the loaded set) so the AI sees the new tools right away without a restart.
OpenAI summary-model selector
Pick a dedicated summarization model for OpenAI-compatible endpoints. Settings now discovers available models from your endpoint (with a refresh button) and lets you choose a separate, usually cheaper, model for conversation summaries — or fall back to your chat model.
Session search across message text
Search your history by what was actually said. Session search now matches against message content, not just session names and summaries.
🐛 Fixes
Chat scrolls to your message on send. Sending a message now snaps the view to the bottom immediately, instead of only following along if you were already near the bottom.
Cleaner history trimming at turn boundaries. When old messages are dropped to fit the context budget, the kept window now always starts on a complete user turn — tool-call/result pairs are never left orphaned, which could confuse the model.
Windows app icon. The Windows build now embeds the Hobbes icon into the executable.
Hobbes v0.10.0 — Smart Context, AI Timers & Message Queue
✨ New Features
Smart context handling
Big tool results stay intact on large-context models. Previously, fetching a lot of data (e.g. a full inbox of emails) could get silently condensed mid-turn even on Gemini/Claude with their huge context windows — which sometimes led the model to "fill in the gaps" with plausible-but-wrong details. Tool results in the current turn are now kept in full, so the model always sees the real data it's working with.
Per-turn budgeting is now provider-aware. Hobbes sizes tool results against each model's actual context window instead of a one-size-fits-all cap. Small local models still get aggressive compression (so they don't overload); large models keep far more in context.
Knowledge-preserving history summaries. When older tool results no longer fit the budget, Hobbes replaces them with a dense summary that preserves the concrete facts (IDs, names, numbers, dates) — instead of chopping the data behind a "fetch the rest" footer. The full data is still retrievable on demand.
Auto-recalibration on "prompt too large". If a provider rejects a prompt as too big, Hobbes trims and retries in-turn, and remembers the real limit so later turns just work — no manual fiddling.
Correct OpenAI context windows. OpenAI-compatible endpoints now resolve their window from the model name (gpt-4.1 → 1M, gpt-5.4 → 1M, gpt-5.5 → 512K, gpt-5 → 400K, gpt-4o → 128K, o-series → 200K), with runtime self-correction for anything unknown.
AI-settable timers & reminders
The assistant can now set timers and reminders that notify you (or prompt a follow-up turn) when they fire, with a live pending-timer indicator above the chat bar and an auto-dismissing toast. Window-focus behavior is opt-in via a Behavior setting.
Send-while-streaming message queue
You can now queue a message while a turn is still streaming instead of waiting — it's sent automatically once the current turn completes.
🐛 Fixes
Windows scrollbars cleaned up. The chat input and side panels no longer show chunky, always-visible scrollbars with arrow buttons on Windows/Linux — the input grows cleanly and then scrolls with a thin scrollbar, matching the macOS look. (macOS overlay scrollbars are untouched.)
Chat bar spacing. Added breathing room below the chat bar and fixed padding that was being clipped by the app shell's overflow.
Hobbes v0.9.62 — Pricing Accuracy
🐛 Fixes
OpenAI cost no longer mis-prices version families. Model matching is now version-aware: gpt-5.1 / gpt-5.2 / gpt-5.3 can no longer silently inherit bare gpt-5's cheaper rate. Recognized families (gpt-5, gpt-5.4, gpt-5.5, gpt-4o, o-series) resolve correctly; an unrecognized model reports no cost rather than a wrong one.
🛠 Maintenance
Verified Gemini and Claude pricing tables against current published rates (all correct as of 2026-06-17).
Documented the model pricing tables as manually-maintained and drift-prone (SYSTEM_PATTERNS P-013), with per-table "last verified" dates.
Hobbes v0.9.61 — OpenAI Cost Tracking & Windows Settings Reliability
✨ New Features
OpenAI API cost tracking
Real per-turn USD cost for OpenAI models — input/output token pricing for the gpt-5.5 / 5.4 families, gpt-5 / mini / nano / pro, gpt-4.1, gpt-4o, and the o-series.
Billed only for the real OpenAI API (an API key present and an api.openai.com endpoint). Local / self-hosted / keyless endpoints (Ollama, vLLM, LM Studio, proxies) stay free — no fabricated costs. Unknown models report no cost rather than a wrong one.
🐛 Fixes
Windows settings reliability
Settings now save reliably on Windows.settings.json is written atomically (temp file → fsync → rename) instead of a non-atomic truncate-and-write, eliminating the corruption/race that made settings silently revert to defaults after the first save. Save failures now surface in the UI toast instead of failing silently.
Hobbes v0.9.60 — OpenAI Responses API Support
✨ New Features
OpenAI Responses API (gpt-5 / o-series)
Support for OpenAI's newest models — gpt-5 and o-series reasoning models are served only by the Responses API (/v1/responses), which previously failed with a 404 — only supported in v1/responses error. Hobbes now speaks both protocols.
Automatic routing — Under the default Auto API style, OpenAI's gpt-5/o-series models on api.openai.com use the Responses API; everything else (older OpenAI models, local servers like vLLM/Ollama/LM Studio, OpenRouter) continues to use Chat Completions. No regression to existing setups.
Reasoning summaries — With Thinking Mode enabled, reasoning summaries stream into the "Considering…" bubble.
Robust streaming — Function calling, multi-turn tool history, vision input, and usage reporting are mapped to the Responses item/event model. A final-text fallback ensures models that reason silently before answering still surface their output.
🛠 Maintenance
Added regression tests for Composio auth-error detection and toolkit-name conversion.
Documented the Dioxus use_memo-over-props streaming footgun (SYSTEM_PATTERNS P-012).
Generate images directly in chat — New generate_image tool powered by Gemini image models, available as a virtual MCP server (hobbes-native-image)
Image editing support — Pass a previously generated image as a reference_image to riff on or modify it
Inline image rendering — Generated images display inline in the chat with download controls; file:// paths converted to data URIs for reliable WebView rendering
Image Generation Model selector — Configurable in Settings → Model panel
Per-Provider Context Tuning
Fine-grained context budget controls — New ContextTuningPreset system lets you override chat history length, tool output limits, summary size, entity count, and budget ratios per provider (Gemini, OpenAI-compat, Claude)
Compact tool results — Option to convert JSON tool results to compact markdown for smaller context models
Dynamic tool result budget & pagination — Pass 2 budget allocation system automatically paginates oversized tool results with a HOBBES_PAGE_RESULT continuation tool
Stream & Thinking Improvements
Thinking content visibility — "Considering..." bubble now shows thinking content in real-time (thinking-only messages forwarded to UI)
