Caveman (YC Fall 2026): teardown vs XMarts

In one line: Caveman started as a free open-source add-on that makes coding agents use fewer tokens. It now sells a cloud product that shows what agents cost, finds waste, tests fixes and counts only the savings it can prove [site] [github]. Threat: 2, because it is the same buyer and the cost half of W2, but it is one person, has no named customer, and stops at cost [yc] [site].

1. Fact file

2. Their proposition, reverse engineered

3. Positioning

• Category they claim: "optimization stack for AI agents"; "self-driving software for your LLM stack" [yc] [docs].
• Category a YC partner would put them in: AI cost control and observability for developers (FinOps for agents), with an open-source launch (inference).
• What they want to be known for: the independent referee that proves an agent cost cut did not hurt quality.
• Where they sit on the map: layers L1/L2/L3 = 1/1/2, buyer = size unknown (sells to teams of all sizes; no named customer), developer.

4. Three layers vs ours

5. Same and different

6. Wedge overlap

7. Whitespace (where they stop)

• Layer gap: they stop at "cost and quality of the call". They do not claim to know whether a task finished, landed or was silently dropped [cloud]. To own it we need a named list of silent-failure types detected on a customer's live agents, with counts.
• Buyer gap: everything is built for developers and platform teams. No product for a commercial or finance lead at a non-technical SME [docs] (inference from the paths offered). To own it we need such a buyer saying this pain is theirs. We have no calls on this yet.
• Number gap: their one number is money saved. They do not report work finished per pound spent. Their letter says cost per finished task should become a standard metric [blog], so this gap may close. To own it we need one customer with a measured "accepted work per pound" figure.
• Price gap: none. Free tier and $300 a month are cheap for a small team [pricing].
• Speed gap: none. Changing a base URL is fast [docs].

8. For the YC application

• Proven today: Marton's logs show the pain at first hand: a single launch authorised 156,000 agent turns and used 39% of the weekly Claude budget in 14 hours, it recurred, and the burn guard then silently stopped. The same logs show he built a heartbeat, a "silent costs" meter and about 160 failure records with automatic checks. Against Caveman: that is cost plus whether the work produced anything. Only about 15% of jobs reached main, which is the honest other side.
• Draft answer to "how are you different from Caveman?": DRAFT. "Caveman proves a cost cut kept quality. We watch whether the agents finished any useful work at all, and where it silently stopped. We have run that on our own fleet of about 5,600 jobs." Founders say it their own way.
• Future claim (needs proof): "We tell a team how much accepted work each pound of agent spend bought." Needs one team's live agents watched for 4 weeks, with a before/after on stuck or silently failed work.
• What we understand that they don't: the costly failures are agents that run and produce nothing, not agents that talk too much. proven on our own fleet (Silent Burn, silent-success recurrences); needs proof on any outside team.
• Copyability (M3): easy. They already record failures and retries and tie sessions to merged changes [cloud]. Their letter names cost per finished task as the goal [blog]. Adding stuck and silent-failure detection is a feature on their buyer (inference). What they lack is our named failure classes from a real fleet, which is a data edge but not a hard one. Not hard, so M3 can not score above 3 against Caveman.
• Why now (M4): None found in this teardown. Do not use "tokens got cheaper but bills rose": that is their claim, undated here and not checked.
• Do not claim: that we have customers or a portable product. "Nobody measures whether agents produce" (Caveman measures cost and failed requests). Any savings percentage. "Independent verification" as our own headline: it is their line.

9. Ideas worth taking

1. Four labelled number types (measured, inferred, tested, verified), with savings shown as $0.00 until proven [cloud] [enterprise]. Shows honesty and fits a YC partner who distrusts agent claims.
2. Publish a result you refused to ship, and a case where the fix made things worse [news]. Trust as marketing.
3. Free open-source tool first, paid cloud second [github] [pricing]. Reach before revenue. Worth asking whether any piece of our system could be given away.

10. Unknowns

• Who the founder is and whether he is solo. The YC page shows no names. Look for a YC launch post or his own site.
• Whether any company pays for Cloud. Look for named logos after launch, and ask in the YC batch.
• How "Autopilot" and "agent runs" work. The Cloud page does not describe them; the docs index names Improvements but I found no page on it (UNVERIFIED).
• Funding. None found. Check Crunchbase-style listings later.
• UK or Europe sales. Company is in Amsterdam [yc]; no UK page found.
• Whether routing and harness optimisation ship (marked "soon" and "dev") [pricing].

Sources

• caveman.so [site]
• Pricing [pricing]
• Caveman Cloud [cloud]
• Enterprise [enterprise]
• Compare [site]
• News index, posts Jul to Sep 2026 [news]
• Efficiency and the token economy, 26 Jul 2026 [blog]
• Docs index [docs]
• YC company page [yc]
• GitHub: JuliusBrussee/caveman [github]
• Gfinity, citing 404 Media, 6 Jul 2026 [press]
• 404 Media (seen in search results only) [press]
• 36kr on the open-source launch, 7 Apr 2026 [press]