Finally, an inbox you'll look forward to. Agents sort your messages, draft your replies, and clear the grunt work behind the scenes, all in a client so well-crafted that email feels light, fast, fun.
No reviews yetBe the first to leave a review for Upstream
Wispr Flow: Dictation That Works EverywhereStop typing. Start speaking. 4x faster.
Promoted
The voice-matching per audience is what caught my eye — does it learn over time from how you edit the drafts, or is it mostly set from the initial analysis?
@doganakbulut Right now it’s mostly set from the initial audience analysis. However, the tone of the drafts do change according to the context of the thread.
That said, learning from how you edit drafts is definitely on our roadmap. It’s the natural next step: the more you tweak tone, phrasing, and structure for a given audience, the better Upstream should get at matching that voice over time.
I would add that each time we generate a draft we use your most recent conversations with the recipient in question to generate the reply, so if you tweak the drafts they will naturally converge toward the writing style you use for a given recipient
@jontiret The agentic drafts are definitely one of the strengths of upstream. This, plus the fact you can integrate it other data sources like notion or google drive makes for excellent drafts!
yeah good point @skwph, for now this is google drive and notion but later potentially with 1000s of sources so that agents have full context
Report
the "agents do the grunt work" angle is interesting because most AI email tools still make you manage the AI. curious how the triage actually works in practice, does the agent learn your priorities over time or is it more rule-based out of the box?
One thing we’ve learned is that people have surprisingly different definitions of what important means 😅
That’s why we start with sensible defaults and then let users customize the behavior from there
Report
The human + agent inbox positioning is interesting. Email has a lot of hidden context, so I like that you are not framing this as just another summarizer. How do you decide which actions should stay manual versus agent-assisted?
@xuanlin right now agents help you draft, follow up, or triage the spammy emails. And you can ask in the chat for more stuff. You can always do anything manually obviously
Curious how you would want things to work ideally?
Also, one principle that is important to us is that even if agents prepare a draft for you, sending is always done with your approval
Report
The hardest part of AI email assistants seems less about drafting and more about trust. At what point do users become comfortable sending agent-generated replies with minimal edits, and where do they still prefer to stay fully in control?
@geoffroy_de_cooman1 As soon as you join, Upstream creates a custom writing agent that analyzes writing style, tonal shifts, lowercase usage, abbreviations, greetings, sign offs, and more for each audience. We also adjust the tone of AI drafts depending on the context of the current conversation. On top of that you can even customize the prompt to fine-tune exactly how you want to sound.
@thomas_df Not yet natively. The current workaround is to open your codebase in Claude Code, Codex, or Cursor, let it gather the relevant repo context, and then have it use Upstream via MCP to draft the reply.
@louislecat today I have a Codex automation that drafts me Gmail answers for Sales & Support messages using the codebase as context, this is really useful!
@thomas_df That's a great use case! For now, you can do the same with Upstream MCP, and agents can also have their own service accounts so they can tag you and post private comments for you and your team on email conversations, or simply reply directly on your behalf like what you describe with Gmail
congrats team! I've a quick question. How can I edit the soul or voice of the agent when writing email draft? Can I create multiple profiles, eg. when writing to a team member, a lawyer, a customer, or a VC?
@jelorillon Thanks Jean! You can edit the prompt for your voice by going to Settings > Configure Agents. You'll be able to tweak your global writing style prompt as well as the prompts for each audience. You can also create a custom audience and define your voice for it.
Also, when you sign up, our custom writing agent will review your past emails to identify your audiences (the typical categories of people you interact with) - most likely the categories you mentioned will already be filled out in the "audiences" section during the onboarding :)
The voice-matching per audience is what caught my eye — does it learn over time from how you edit the drafts, or is it mostly set from the initial analysis?
Upstream
@doganakbulut Right now it’s mostly set from the initial audience analysis. However, the tone of the drafts do change according to the context of the thread.
That said, learning from how you edit drafts is definitely on our roadmap. It’s the natural next step: the more you tweak tone, phrasing, and structure for a given audience, the better Upstream should get at matching that voice over time.
Upstream
I would add that each time we generate a draft we use your most recent conversations with the recipient in question to generate the reply, so if you tweak the drafts they will naturally converge toward the writing style you use for a given recipient
Upstream
@jontiret The agentic drafts are definitely one of the strengths of upstream. This, plus the fact you can integrate it other data sources like notion or google drive makes for excellent drafts!
Upstream
yeah good point @skwph, for now this is google drive and notion but later potentially with 1000s of sources so that agents have full context
the "agents do the grunt work" angle is interesting because most AI email tools still make you manage the AI. curious how the triage actually works in practice, does the agent learn your priorities over time or is it more rule-based out of the box?
Upstream
Hey Raj, a couple things come out of the box:
1 agent identifies everything that needs a reply and puts it in 1 place
1 agent identifies the noise / spammy emails and moves them to a less intrusive place
1 agent identifies when people don't reply to you and brings it back at the exact right time
And from there you can customize and add more triaging options
Upstream
One thing we’ve learned is that people have surprisingly different definitions of what important means 😅
That’s why we start with sensible defaults and then let users customize the behavior from there
The human + agent inbox positioning is interesting. Email has a lot of hidden context, so I like that you are not framing this as just another summarizer. How do you decide which actions should stay manual versus agent-assisted?
Upstream
@xuanlin right now agents help you draft, follow up, or triage the spammy emails. And you can ask in the chat for more stuff. You can always do anything manually obviously
Curious how you would want things to work ideally?
Upstream
Also, one principle that is important to us is that even if agents prepare a draft for you, sending is always done with your approval
The hardest part of AI email assistants seems less about drafting and more about trust. At what point do users become comfortable sending agent-generated replies with minimal edits, and where do they still prefer to stay fully in control?
Upstream
@yagnaveena the user maintains control in both scenarios. You can edit agent generated drafts or completely write them from scratch.
Love your promise! Out of curiosity: every other email tool say they sound like me these days, what makes you claim you are better here?
Upstream
@geoffroy_de_cooman1 As soon as you join, Upstream creates a custom writing agent that analyzes writing style, tonal shifts, lowercase usage, abbreviations, greetings, sign offs, and more for each audience. We also adjust the tone of AI drafts depending on the context of the current conversation. On top of that you can even customize the prompt to fine-tune exactly how you want to sound.
Upstream
I couldn't have said it better! :)
Twenty
Congrats Louis, Jonathan, and team! Can you plug your codebase/github repo as context?
Upstream
@thomas_df Not yet natively. The current workaround is to open your codebase in Claude Code, Codex, or Cursor, let it gather the relevant repo context, and then have it use Upstream via MCP to draft the reply.
What’s the use case you have in mind?
Twenty
@louislecat today I have a Codex automation that drafts me Gmail answers for Sales & Support messages using the codebase as context, this is really useful!
Upstream
@thomas_df That's a great use case! For now, you can do the same with Upstream MCP, and agents can also have their own service accounts so they can tag you and post private comments for you and your team on email conversations, or simply reply directly on your behalf like what you describe with Gmail
Upstream
@thomas_df oh the codex thing seems really cool!
Upstream
@thomas_df Thank you Thomas!
Upstream
@thomas_df thanks a lot! Appreciate it
Sales Compensation Hub
congrats team!
I've a quick question. How can I edit the soul or voice of the agent when writing email draft? Can I create multiple profiles, eg. when writing to a team member, a lawyer, a customer, or a VC?
Upstream
@jelorillon Thanks Jean! You can edit the prompt for your voice by going to Settings > Configure Agents. You'll be able to tweak your global writing style prompt as well as the prompts for each audience. You can also create a custom audience and define your voice for it.
Just curious what are your most common audiences?
Upstream
Also, when you sign up, our custom writing agent will review your past emails to identify your audiences (the typical categories of people you interact with) - most likely the categories you mentioned will already be filled out in the "audiences" section during the onboarding :)