Roadmap
Help Center
Changelog

Help me improve BoltAI

For a bug report, please include the following information:

  • Your OS version.

  • Your BoltAI app version.

  • Are you a Setapp user.

  • The AI provider & model you’re using.

  • Steps to reproduce the issue.

  • The error message

Planned

BoltAI issues due to outdated documentation

OS: MacOS Tahoe 26.5 BoltAI V2 I am having several issues with BoltAI because the documentation appears to be outdated. Google web search does not seem to work given that it no longer supports searching the entire web. Next to that, I can't get any models in BoltAI to perform file operations like Codex does, even though the documentation mentions support for both a Shell and an AppleScript plugin. Neither plugins seem to show up anywhere, no matter which model, cloud or local. Also, the documentation mentions setting a path variable in the settings pane, but I can't find that anywhere either.

Andy De Meyer about 18 hours ago

1
🐛

Bug Reports

MCP tool input schemas are stripped of type/enum/default — some properties sent as empty {}

When forwarding MCP tools to Claude/Anthropic, BoltAI appears to apply a lossy transformation to each tool's input_schema, dropping type information that the MCP server provides: Several properties are emitted as empty schemas with no type, e.g. "query": {}, "continuation": {}, "channel": {}, "playlist": {}. Properties that have enum/default in the source schema lose them — e.g. format is sent as {"type": "string"} instead of {"type": "string", "enum": ["json","text"], "default": "text"}. This is valid JSON Schema ({} matches any type) and is accepted by the API, so it's not a hard error — but it's a quality regression. With no type, the model has to guess the expected argument type, which reduces tool-call reliability (malformed or wrong-typed arguments). Dropped enum/default values remove guidance the model would otherwise use. Confirmed by comparison: another client (not named) hitting the same MCP server forwards the full schemas intact — typed properties plus enum and default values — while BoltAI's version is stripped down. This points to a transformation in BoltAI's MCP→Anthropic schema handling rather than the MCP server's output. Request: forward the MCP server's tool inputSchema unchanged — preserve each property's type, enum, and default.

David 1 day ago

🐛

Bug Reports

in BoltAI - MCP tool names use single-underscore mcp_ prefix → Anthropic rejects Claude (OAuth) requests with HTTP 400

When a Claude/Anthropic request includes MCP tools, BoltAI names them mcp (single underscore, e.g. mcpsearch_youtube_z50gm8). Anthropic classifies any tool whose name starts with mcp (single underscore) as a third-party MCP tool and bills it to "extra usage," returning: 400 invalid_request_error: "Third-party apps now draw from your extra usage, not your plan limits…" On accounts with extra usage disabled, every MCP-tool request fails. Claude Code's own convention is a double underscore — mcp — which Anthropic accepts as first-party. Verified fix: renaming mcpsearch_youtube_z50gm8 → mcp__search_youtube_z50gm8 (only change: one extra underscore) turns the 400 into a 200. Requests without tools, or with tools that don't use the mcp prefix, already work. Request: name MCP tools with the mcp__ double-underscore convention. Tried it with https://github.com/router-for-me/CLIProxyAPI and https://github.com/automazeio/vibeproxy (and used in AI Providers - Anthropic endpoint and pointed it to CLIProxyAPI)

David 1 day ago

4
💻

Feature Requests

Cross-Chat References

Allow referencing or including context from previous chats in a new conversation. What I'd like: Ability to attach/link a previous chat (or multiple chats) as context to a new one Ideally searchable chat history so finding the right one is easy attach/link selectable or all history from Projects Why it matters: Work often spans multiple sessions — being able to pull in context from a past conversation without copy-pasting saves time and avoids losing important decisions/outcomes.

David 10 days ago

4
💻

Feature Requests

Tabbed Parallel Chats

Add a tab bar to the chat interface (at the top) so multiple conversations can run simultaneously without switching between separate windows. It would show only (active?) chats at the top… something like in the screen shoot.. Each tab runs independently (LLM calls don't block each other) Easy switching between tabs, ideally with keyboard shortcuts (cmd+left/right)

David 10 days ago

💻

Feature Requests

BUG: UI hangs when running mcp_add_memory_tctnzi tool

OS: macOS 26.4.1 Version: BoltAI 2.12.1 (build 68) - Setapp LLM: Anthropic Claude Sonnet 4.6 I instructed the LLM to commit something in particular to memory. It appears to have called the mem0-memory-mcp server installed. Although it’s not clear to me if that’s what Bolt tried to use, or whether it’s using the built-in Bolt memory tool. I can see in the UI what you see in attached screenshot. The conversation hang after it tried to use this memory tool. I had to manually terminate the hung prompt response. I can see that the number of memories in the Bolt Memory tool didn’t increase. So if that is the memory system it’s utilising in this chat, the hung attempt to add a memory did not get so far as actually adding a memory.

Inspired Earth 18 days ago

2
💻

Feature Requests

Google Deep Research models not working

When connecting through “Google AI” AI Providers and when selecting any Deep Research model there is an error: Error This model only supports Interactions API. I have all MCPs and plugins disabled… all other Gemini models work. https://ai.google.dev/gemini-api/docs/interactions/deep-research i guess you would need to implement an additional api endpoint for this to work?

David 18 days ago

💻

Feature Requests

Choice of local models for STT

Not having a choice in speech-to-text models for the dictation feature is very limiting. The two pre‑selected models are not great. Silences are always replaced with lots of random words. And the user doesn't have the ability to pick from a whole variety of excellent speech-to-text models now available. Look at something like Mac Whisper and the flows that exist there.

Thinkun 19 days ago

4
💻

Feature Requests

Built a Hermes Agent gateway plugin for BoltAI — full markdown + slash commands

Hey, I've been using BoltAI v2 with Hermes Agent (https://github.com/NousResearch/hermes-agent) and hit two annoyances with the official OpenAI-compatible gateway: Markdown gets flattened — headers, tables, code blocks all come through as plain text Hermes slash commands (/help, /status, /new, /model, etc.) aren't intercepted, so they get sent to the LLM instead of running locally and therefore they dont work properly as AI agent cant work with some directly… So I built a small hermes drop-in plugin that fixes both. It's a separate gateway adapter (subclasses the official one, doesn't fork it), runs on its own port, and lives here: https://github.com/Dworf/boltai-hermes-gateway Install: hermes plugins install Dworf/boltai-hermes-gateway --enable Then in BoltAI just follow the official Hermes setup guide (https://help.boltai.com/en/help/articles/9536908-how-to-use-hermes-agent-in-boltai) but point it at port 8643 instead of 8642. Full markdown renders correctly, slash commands work as expected, streaming is preserved. Should also work with any other OpenAI-compatible client that supports markdown + streaming (TypingMind, OpenWebUI, LibreChat, etc.) — untested though (and yeah BoltAI has no competition in this space :) ) Question for the BoltAI team / community: if there's interest, I'm happy to implement a commands discovery endpoint on the plugin side so that — if BoltAI ever wants to add it on the client side — you could get interactive slash command suggestions, autocomplete, and /-menu UX natively. Let me know if that's something worth building toward. Should be easy from the plugin side… Thanks!

David 23 days ago

2
💻

Feature Requests

In Progress

Model popover fetching models for a very long time

I have a couple of models on openrouter.ai and a few via Ollama on a PC in my local network set up. The local PC is turned off most of the time. Recently this popover where you can switch the current model started fetching models for a long time (30s+) and I had to close and re-open BoltAI 2 to work around it. My internet access at that time is working as usual, I can also open my openrouter.ai profile without any problems. Let me know how I can submit any logs to help debug this.

zuKoma091 23 days ago

2
🐛

Bug Reports

model selection of model capability based on output (text, image, video,...)

I really like the new model selection… and love the favorites. but when working with model collections like (image generation, video generation,…) it would be good to have filter option to only show models that can do image generation or video generation,…. i would be happy to just tag each model if it supports for example image generation… i think you could pull this data as well from openrouter for example…

David about 1 month ago

💻

Feature Requests