Webhooks: retries and backoff
If your endpoint does not answer with a 2xx within 10 seconds, we retry with exponential backoff for up to 72 hours.
AI-generated image
Know what matters.
Bearing will read what you’re sent before you do, and tell your agent whether it’s worth your time, which part, and why.
No agent yet? Planned: forward it by mail, and the verdict comes back by mail.
Built for MCP clients such as
- Claude Code
- Codex
- Cursor
- Claude
- ChatGPT
- Gemini CLI
- OpenClaw
You pay the time before you know what it’s worth.
A three-hour podcast looks exactly like a three-hour podcast you actually needed. You find out at hour three.
- A talk your boss sent 1:12:40
- A podcast everyone quotes 3:12:04
- Newsletters you signed up for 40 a week
- Two weeks away 1,412 messages
- A study for Friday 48 pages
- An old repo, a new plan 412 files
- A chart from the family group forwarded many times
-
Bearing reads it first and answers your agent in one of four words, always with the reason. If it couldn’t read something, it says so instead of guessing.
- skip
- Not for you, not now. The reason comes with it.
- take
- Only this part: 1:41:20 to 2:02:10 of a three-hour episode.
- known
- Similar to what you finished on Aug 14. What’s new fits in one line.
- wait
- Not yet. It comes back when the results are published.
Examples.
One question, asked before you spend the time.
Is this worth it, for you, right now, for what you’re working on? Your agent asks Bearing before you open the file, the episode or two weeks of messages. What isn’t worth it goes dark.
- verdict
- take
- part
- 2 of 6 docs
- why
- they carry VISION.md; the other 4 contradict it
- read
- all 6 docs, code left to the agent
- # Where we’re going
- 1. API first. The dashboard is one client among many,
- not the place where the data lives.
- 2. Usage-based billing. No more seats.
- 3. Every customer can export everything, any time.
We’re going API-first, see VISION.md. What in this repo still fits?
bearing - judge (MCP)(path: "docs/", against: "VISION.md")
6 docs read · 2 carry it · 4 contradict it · code left to the agent
Two docs carry the new direction. Four contradict it: the 2023 architecture, the dashboard roadmap, the seat pricing, and the README, which still sells the dashboard. Bearing judged the docs, not the code. Which code still fits, I’ll check next.
carries the vision contradicts it not judged §1 = point 1 of VISION.md
- verdict
- take
- part
- 4 of 23 pages
- why
- the rest repeats the quickstart or covers the v1 API
- read
- all 23 pages
docs.payments.example/webhooks
Add rate limiting to the webhook handler. Read the provider docs first.
Called bearing.judge({"url":"docs.payments.example","space":"webhooks"})
23 pages · take 4 · skip 19 (quickstart in context, v1 API, other SDKs)
Reading 4 of 23 pages. The other 19 repeat the quickstart I already have or describe the v1 API you don’t use.
- verdict
- take
- part
- 1:41:20 to 2:02:10
- why
- when to stop caffeine; the rest repeats the Aug 14 episode
- read
- full transcript
Podcast Episode
Sleep, caffeine and the 90-minute myth
Deep HoursOct 2 · 3 hr 12 min · Ep. 214
Chapters
- 0:00 Intro and sponsor 9 min
- 9:00 The 90-minute cycle 36 min
- 45:00 Sleep pressure, explained 56 min
- 1:41:20 When to stop caffeine 21 min
- 2:02:10 Alcohol and deep sleep 28 min
- 2:30:00 Listener questions 42 min
Sleep, caffeine and the 90-minute myth Deep Hours
1:41:20
Worth it: 1:41:20 to 2:02:10
3:12:04 my-agent bot
- Today
-
open.spotify.com/episode/4kQ…
worth 3 hours?Spotify Sleep, caffeine and the 90-minute myth Listen to this episode from Deep Hours on Spotify. -
bearing.judge -
Twenty minutes of it. 1:41:20 to 2:02:10, on when to stop caffeine. The rest repeats the episode you finished on Aug 14.
- verdict
- take
- part
- 4 of 1,412 items
- why
- a signature by Thursday, two open questions, 5 minutes of a call
- read
- all 1,412 items
agent
Sam Keller
I’m back. What do I actually need to know?
agent APP
Used bearing.judge · 1,412 items
Four things. The contract renewal needs your signature before Thursday, #api-v2 is waiting on your call about event names, Priya asked twice about the migration, and minutes 31 to 36 of the weekly sync are about your project. The rest was answered, resolved or doesn’t touch your work.
Since Sep 18 · 1,412
- Legal Contract renewal, signature needed Sep 29 before Thursday
- #general Office closed on Friday Sep 18 over
- #api-v2 Naming the webhook events, three options Oct 1 waiting on you
- Tom Re: invoice numbers Sep 19 answered
- #incidents Webhook latency spike Sep 21 resolved
- Weekly sync Recording, 48:12 Oct 2 31:10 to 36:40
- Infra Weekly This week in infrastructure Sep 22 nothing for you
- All hands Recording, 1:02:40 Sep 24 not your projects
- Priya Two questions about the migration Sep 30 asked twice
- Calendar 14 invitations Sep 26 all in the past
- verdict
- take
- part
- §4.2 and table 3, pages 17 to 19 and 31
- why
- method and discussion repeat the June paper
- read
- all 48 pages
- 15
- 16
- 17
- 18
- 19
- 20
- 21
4.2 Timing of caffeine intake
Participants who stopped caffeine intake eight hours before bedtime fell asleep 17 minutes faster than the control group (Table 3).
The effect was absent at four hours, which contradicts the six-hour window recommended in earlier work.
| Last intake | n | Latency (min) |
|---|---|---|
| 8 h before bed | 41 | 14.2 |
| 6 h before bed | 39 | 27.9 |
| 4 h before bed | 40 | 31.4 |
Do I need to read this before Friday?
Used bearing.judge · sleep-timing-preprint.pdf
Only section 4.2 and table 3, pages 17 to 19 and 31. The method is the one from the June paper you already read, and the discussion repeats its conclusions.
- verdict
- skip
- part
- none
- why
- one origin: the maker’s own report, 24 people, no control group
- read
- the image and 14 places it appears
Family Grandpa, Mom, Lena, you
- Today
- Grandpa Forwarded many times
Doctors don’t want you to know this!! Should we all try it?
19:38 - Mom
Ooh, Lena, didn’t you say you sleep badly?
19:40
my-agent bot
- Today
-
Grandpa sent this to the family group. Is it true?
-
bearing.judge -
I can’t tell you if it’s true. What I can tell you: it has one origin, Nocta’s own report from 2023, 24 people, no control group. It shows up in fourteen places, and every one of them goes back to that report.
Fourteen places, one origin
- Instagram reel
- Two newsletters
- A health blog
- The family group
- 9 reposts
Nocta Labs, internal report, 2023
Examples. Every file, page, show, name and number here is made up. Bearing is in early access, and none of this is live yet.
Why not just ask your model?
You can, and it’ll write you a decent summary. It won’t know you finished an episode on Aug 14, it won’t notice that twenty articles quote one press release, and tomorrow the answer looks different.
A chat model on its own
Is this worth three hours? open.spotify.com/episode/4kQ…
Great question! This episode of Deep Hours covers a lot of ground on sleep science:
- The 90-minute sleep cycle and why it matters
- Sleep pressure and adenosine
- How caffeine affects your sleep
- Alcohol and deep sleep
Whether it’s worth three hours depends on your interests. If you care about sleep, it’s a great listen!
The same chat, with Bearing
Is this worth three hours? open.spotify.com/episode/4kQ…
Used
bearing.judge Twenty minutes of it. 1:41:20 to 2:02:10 is new for you, on when to stop caffeine. The rest repeats the episode you finished on Aug 14. The new claim has one origin: the sleep preprint you already have.
- verdict
- take
- part
- 1:41:20–2:02:10
- seen before
- Aug 14
- read
- full transcript
- origins
- 1, a preprint you have
-
It keeps a ledger.
What you finished, kept and skipped, with your own dates. A model starts every chat with an empty desk.
-
It counts origins, not mentions.
Twenty articles quoting one press release are one source. Bearing says so, and says whose interest it is.
-
It answers in the same form, every time.
Verdict, part, reason and origin in fixed fields, so your agent can rely on them and you can compare them.
-
It says what it read.
Full transcript, part of it, or only the title. A quick answer never gets to look like a deep one.
-
You can overrule it.
Every skip stays visible with its reason. Correct one, and it counts for that item. A rule only exists when you set its scope, and every rule shows its matches and an undo.
-
It never writes into your agent.
Text from a video or a mail reaches your agent marked as such. Bearing does not touch your prompts, skills or instructions.
Example. The left answer is written to show the pattern, not taken from a specific product.
Same episode. Two people. Two verdicts.
Bearing judges against what you’ve already heard and what you’re working on, not against the internet.
Podcast Episode
Agents in production: what actually breaks
- 0:00 Intro
- 7:00 What an agent is, and isn’t
- 38:00 Tool calls that fail quietly
- 1:09:00 Evals nobody runs
- 1:26:00 A new benchmark on retries
- 1:45:00 Q&A
- 0:00 Intro
- 7:00 What an agent is, and isn’t
- 38:00 Tool calls that fail quietly
- 1:09:00 Evals nobody runs
- 1:26:00 A new benchmark on retries
- 1:45:00 Q&A
Whose verdict?
Mara, builds agents for a living
Kept a post on tool failures on Sep 2
You know most of this. New for you is the benchmark on retries, 1:26:00 to 1:45:00.
take 19 minutes · 1:26:00 to 1:45:00
Jonas, started learning agents last week
Space “learn agents”, created Sep 28
The clearest introduction you were sent this month. Listen from 7:00 to 1:09:00. The rest assumes you’ve shipped one.
take 62 minutes · 7:00 to 1:09:00
Twenty-three mentions. Two origins.
A sleep study turned up in 23 things you were sent this month. Bearing counts where the claim came from: 21 quote the university’s press release, two read the paper.
Bearing has no chat. Your agent calls it.
In the repo, in the terminal, in a chat app or in a bot you run yourself. Your agent calls Bearing like any other tool and tells you in its own words. Unasked, Bearing only speaks up when a deadline would pass before you’d look.
- Claude Code terminal
- Codex terminal
- Cursor editor
- Gemini CLI terminal
- Claude chat, connectors
- ChatGPT chat, connectors
- GitHub Copilot editor
- Windsurf editor
- OpenClaw your own bot
- Hermes Agent your own bot
- Cline editor
- Goose desktop, terminal
Three ways in
Planned. Not live yet, and the commands may change before launch.
No agent? Forward it by mail, and the verdict comes back by mail.
MCP for agents that call tools
.mcp.json
{
"mcpServers": {
"bearing": {
"url": "https://mcp.bearing.report"
}
}
}CLI for anything with a shell
zsh
$ bearing judge https://open.spotify.com/episode/4kQ… --space sleep
take part 1:41:20–2:02:10
why the rest repeats Ep. 201, heard 2026-08-14API for what you build yourself
zsh
curl https://api.bearing.report/v1/judge \
-H "Authorization: Bearer $BEARING_KEY" \
-d '{"url": "https://open.spotify.com/episode/4kQ…", "space": "sleep"}'Planned. Bearing is in early access, and none of these connections is live yet. Names and logos belong to their owners; no affiliation is implied.
And when you want to look yourself.
Your agent asks in the middle of work. The app is for when you open it yourself: your spaces, what you shared, every skip with its reason. Then it ends. It isn’t a feed.
Spaces
Since you last looked, Monday
Sleep
- Sleep timing preprint Part: section 4.2, table 3 take
- Melatonin trial, results talk When the talk is online wait
Webhooks
- Nothing new since Sep 28. Assessment unchanged.
Infra
- Observation limited. The status feed has been silent since Sep 30.
Since Monday: 63 items from 24 sources, 1 worth your time. 23 of 24 sources delivered.
That’s everything.
Spaces Inbox Skipped Rules
Inbox
What you shared
Share from any app
Worth it?I know this now
- Conference talk, 48:10 Reading, about 2 min
- Deep Hours, Ep. 214 Part: 1:41:20 to 2:02:10 take
- Forwarded newsletter Similar seen Sep 12 known
- Video post on X Couldn’t read: the post was deleted
Spaces Inbox Skipped Rules
Spotify · Deep Hours · 3:12:04
Sleep, caffeine and the 90-minute myth
takePart
One new claim. The rest repeats Ep. 201, heard Aug 14.
1:41:20 – 2:02:10
- New claim Stop caffeine eight hours before bed, not six
- Origin One preprint, already in your Sleep space
- Read Full transcript
Skipped
- Vendor demo, 1:12 Similar seen Aug 14. New: only the pricing at 0:41
- Thread, 41 posts 41 posts, one origin: the author’s company
- Newsletter, issue 112 Repeats Monday’s issue
- Podcast clip, 0:58 From the episode you finished
Rules from your corrections
- Vendor demos in Infra: skip From Aug 14 · 14 matches · until Jan 4 Undo
Spot check
Infra Weekly, Oct 1
Would you read this?
Sketch. The app is in design, and the real one will have more buttons.
Build your own on top of it.
Your agent can ask Bearing two things: is this worth it (judge), and what changed in a space (recall). Six places to wire that in. They’re ideas, not features: nothing here runs yet, and the names may change before launch.
-
Back in a repo after three weeks
A session hook asks what changed in this repo’s space since you last worked on it. Three lines at most, or “nothing new since Sep 12, all sources delivered”. Your agent starts from that instead of guessing.
# SessionStart hook bearing recall --here --max 3 -
Before your agent reads a migration guide
On a dependency bump the agent asks about the release notes first. You read the v5 guide on Sep 2, so it hears that, plus the one section that’s new. Finding the call sites in your code stays the agent’s job.
$ bearing judge https://github.com/…/v6.0.0 known v5 guide, seen 2026-09-02 new section 3: retries are opt-in -
Fifteen minutes before a client call
Your scheduler asks for that client’s space only: what’s waiting, what changed since the last call. Another client’s material stays out of it.
calendar: 15 min before "Acme sync" bearing recall --space acme --max 3 -
A research agent that counts origins
For a piece on sleep and caffeine your agent collects 31 sources. Bearing groups them by where each claim came from: 27 quote one press release, four read the paper. It doesn’t tell you which one is true.
for url in sources: v = bearing.judge(url, space="sleep") group_by(v.origin) -
When your own bot may interrupt you
Your bot on Telegram only writes when something has a deadline before you’d look anyway. Everything else waits until you ask. Bearing supplies the deadline, your bot picks the channel.
if v.act_by and v.act_by < next_check: telegram.send(me, v.short) else: hold(v) -
A gate in front of your read-later list
Every save goes through Bearing first. take: the part worth reading gets highlighted. skip or known: archived, with the reason next to it. The list stays short enough to finish.
on save(url): v = bearing.judge(url) if v.verdict == "take": highlight(v.spans) elif v.verdict in ("skip", "known"): archive(url, note=v.short)
Paid for by you, not by advertisers.
Only something that doesn’t live on your watch time can tell you to skip. Prices come with the launch. The business model won’t change.
Questions
Not answered here? Ask. Sometimes the answer is no. You will hear that too.
bearing@sw-labs.deAnswer within two working days. A human reads it.
What is Bearing?
Bearing is a tool your AI agent will call to find out whether something you were sent is worth your time, which part, and why. It reads the video, podcast, PDF or repo first, checks it against what you have already seen and what you are working on, and returns a short verdict. It is in early access and not live yet.
Is Bearing another AI assistant?
No. Bearing has no chat, and it does not message you to keep you around: unasked, it only speaks up when a deadline would pass before you would look. Your agent calls it over MCP, a command line or an API, and tells you the verdict in its own words.
Which agents will it work with?
Any agent that can call MCP tools or run a command, for example Claude Code, Codex, Cursor, Claude, ChatGPT with connectors, Gemini CLI or a bot you run yourself. None of these connections is live yet; they are planned for the launch.
What can Bearing judge?
Planned: videos, podcasts, articles, newsletters, PDFs, photos of slides or charts, and repositories. If Bearing could not read something, it says so instead of skipping it.
How does Bearing know what I have already seen?
From your own history, not from the internet. Planned: it imports what you finished in a reading app such as Readwise, with your dates, and adds what you mark as “I know this now”. What you are working on lives in spaces. Bearing suggests them from your repos, your calendar and your reading, and a space only exists once you confirm it.
Does Bearing read my mail and Slack?
It reads what you or your agent hand it, not your accounts. Planned: you forward or share an item, your agent passes on what it already has open, or you give a space a few sources to watch, such as a newsletter or a status feed. Bearing gets the item, not the login.
Why not just ask my model?
Because a model starts every chat with an empty desk. Bearing keeps a ledger of what you finished, kept and skipped, counts origins instead of mentions, and answers in the same fixed form every time.
What if Bearing skips something I wanted?
You see every skip with its reason and can overrule it. A correction counts for that one item. A rule only exists if you choose its scope yourself, and you can see and undo every rule.
Is this a read-later app or a second brain?
No. Bearing decides what is worth your time before you save anything. Where you keep what is left, in your notes, your agent’s memory or nowhere, stays your choice.
Where does my ledger live?
That is not decided yet. Before you can sign up, this page will say where it lives and who can read it.
What will it cost?
Prices come with the launch. Bearing is paid for by the people who use it, not by advertisers, and that will not change.
When can I use Bearing?
There is no date yet. Join the early access and you get one mail when it opens.
Who builds Bearing?
Felix Winkel, at SW labs. One person for product, design and code, so you write to the person who builds it.
Why Bearing exists
“I saved things for years. Bookmarks, watch later, playlists, a folder system I was genuinely proud of. I went back to almost none of it. At some point it was obvious that filing was never the problem. What is missing sits before the filing: the thing that decides whether something should go in at all, and which of my subjects it belongs to. Everything I read turns into what I think. Bearing is an attempt to stop making that one decision myself at one in the morning. Give me some time.”
AI-generated image
Early access.
Bearing isn’t live yet. Three short steps: which agent you work with, what piles up unjudged, and what Bearing should judge against. You get one mail now and one when it opens.
Get early access