Centauri AI (YC Winter 2024): teardown vs XMarts

In one line: Centauri AI started as a data tool for structured finance and now also sells coSlash, a free local viewer plus paid team space that shows all your coding-agent sessions on one board [site] [yc]. Threat: 2, because coSlash sits on the same buyer and problem as W2 (seeing what your coding agents are doing), but it is new, tiny and only watches; it does not run or check the work [site] [docs].

1. Fact file

Their traction numbers (they publish none). What GitHub shows: 17 stars, 3 forks, 1,085 commits, and release asset downloads that add up to 102 across seven visible releases (v0.1.0 to v0.2.1; the first release candidate was not visible) [docs]. Download counts exclude Homebrew installs (inference).

2. Their proposition, reverse engineered

3. Positioning

• Category they claim: "The attention layer for coding agents." [docs]
• Category a YC partner would put them in: developer tools for agent session management (inference). A second, older finance data company sits beside it on the same site [site].
• What they want to be known for: the place where a team sees every coding agent and keeps what each one learned.
• Where they sit on the map: L1/L2/L3 = 2/2/1, buyer = sme, developer.

4. Three layers vs ours

5. Same and different

6. Wedge overlap

7. Whitespace (where they stop)

• Layer gap: they watch and hand off; they do not run the agents or check the output [site] [docs]. To own it we need the full loop shown on one team's work: agents run, a different AI checks, and a number moves.
• Number gap: no public result for coSlash, only activity (sessions, status) [site]. To own it we need one before/after number such as accepted changes per week, or rework rate, for a real team.
• Buyer gap: they serve the individual developer and small team [site]. They do not address the lead who must trust unattended work. To own it we need that lead saying they have the pain (no call yet).
• Price gap: none. $15 to $29 a month, with a free local tool [site]. We can not undercut it.
• Speed gap: none. One command install [site]. We can not beat first value in minutes.

8. For the YC application

• Proven today: Marton's logs show his system has run about 5,600 agent jobs with a heartbeat that reports what is stuck, what silently stopped and what it costs, and about 160 written failure records that each added an automatic check. Centauri's public pages show a session viewer with status, context and handoff, and no published result [site].
• Draft answer to "how are you different from Centauri AI?": DRAFT. "Centauri shows you your coding sessions on one board. Our system runs the agents around the clock and catches the failures nobody saw, like silent stops and runaway cost." Founders say it their own way.
• Future claim (needs proof): "Teams trust unattended agent work because we show what failed and what it cost." Needs a second team's agents run for 4 weeks with a named before/after number (for example failures caught before merge).
• What we understand that they don't: the pain is not seeing sessions, it is trusting work nobody watched; the named failure classes come from five months of running it. needs proof on a customer; Marton's logs show the failure records on his own system.
• Copyability (M3): easy. They already read transcripts, track status and ship a release about every week (v0.1.0 on 30 Sep, v0.2.1 on 12 Oct 2026) [docs]. Alerts for stuck or failed sessions would be a feature that fits their buyer (inference). What they lack is a queue, a gate and unattended runs of their own, and their "no daemon" stance points away from that [docs]. Evidence is thin, so this is not hard.
• Why now (M4): None found in this teardown. Do not use "more teams run parallel coding agents": that is general AI progress.
• Do not claim: "nobody watches coding agents" (they do, and they ship weekly). "Silent-failure detection" as unique (UNVERIFIED that they lack it). Anything about their traction beyond the GitHub numbers.

9. Ideas worth taking

1. A free, local, read-only tool as the door, with a paid team space behind it. It removes trust and security objections at install.
2. Handoff as plain Markdown that works across agents. It avoids lock-in and is easy to explain.
3. Watch the pivot as a signal: a YC team with finance-data roots moved to coding-agent supervision in 2026 (inference from the repo date and the YC page). Evidence that the "managing many agents" pain is felt.

10. Unknowns

• Whether the finance product still has paying customers, or coSlash is now the main bet. Look for customer pages, a trust centre, and the next batch of YC profile edits.
• Real usage of coSlash. Check later release download counts, GitHub stars, and any launch post (none found on Hacker News, Product Hunt or Reddit in one search).
• Funding. The only figure (about $500k) comes from third-party listings that disagree on round type. Ask in conversation or check a filing.
• Whether Hub has paying teams. Look for case studies or named customers on coslash.io.
• Whether they sell in the UK or Europe. Look at customer pages and contact details.
• The two Substack posts by Milan Shen on agent failures were not readable. They may show how they think about W2-type failures.

Sources

• Centauri AI company site [site]
• coSlash home page [site]
• YC company page [yc]
• YC jobs page: none posted [jobs]
• YC launch post: modernizing ETL for finance [yc]
• coslash GitHub repo, read 2026-10-19 [docs]
• GitHub API: repo data and releases, read 2026-10-19 [docs]
• centauri-ai GitHub org [docs]
• Third-party summaries seen in a search only, not opened: CB Insights, VCBacked, Startup Intros, 1 Feb 2024 [press]