XMarts room

Spring 2026

Linzumi

Team chat to command all your company's coding agents

Visit website

What they do

Team chat where engineers direct and review many coding agents

How close to us

4, very close

Chat for directing coding agents with decisions kept as shared source of truth and checked work; team size not stated.

Wedges

  • W1 Trusted 24/7 coding for small startup teams direct
  • W2 Seeing whether your AI agents actually produce adjacent
  • W3 Company memory from every call, for consultancies adjacent
  • W6 A tool for product managers adjacent
  • W8 The other 90% of an AI skill adjacent
L1 Company brain (knows the organisation) 1, light
L2 Agent team (runs agents, tools, approvals) 2, clear
L3 Does and learns (finishes, checks, measures, updates the brain) 1, light

Who buys

Firm size
sme
Buyer type
developer
Role
head of engineering or founder of a small engineering team
UK
not known
Their own description (from YC)

Linzumi is the chat your engineering team uses to direct dozens of AI coding agents.

Kick off work, review it, ship it. All from chat.

See every agent's work. Unblock the agents that need you from your laptop, your phone, the coffee line. Your teammates see it too.

Every decision your team makes compiles into a source of truth - what to build, why it changed, how your agents should act. That same source generates new work and verifies completed work.

Deliberative alignment, repurposed from aligning LLMs to aligning companies.

Full teardown

Linzumi (YC Spring 2026): teardown vs XMarts

In one line: Linzumi is a team chat where engineers start, watch, steer and review many coding agents that run on their own computers [site] [yc]. Threat: 2, because it is the same buyer and problem as W1 and has a seed round, but it is 3 people, in beta, with no named customer and one agent maker (Codex) [site] [yc]. Expect a YC partner to name it, because its roadmap word is "decisions compile into a source of truth" [yc].

1. Fact file

Item Finding Source
What they say "A chat where every channel ships code." YC line: "Team chat to command all your company's coding agents." [site] [yc]
Founded, team, funding Founded 2025, 3 people, San Francisco. Backers listed: Y Combinator, Matrix, SV Angel, Decibel, Pioneer Fund, Axiom Partners. Amount not public. One directory dates the seed to Dec 2025 (not confirmed by a second source). [site] [yc] [press]
Founders Sean Grove, CEO. Ex OpenAI (post-training and alignment, model-spec work). Founded OneGraph (YC S18, bought by Netlify, Nov 2021), then Netlify Principal Architect. The YC launch post calls him a 3x YC founder. What YC likely saw: a repeat founder with an exit and the exact "write the spec, align the model" background (inference). [yc] [press]
Customers None named. No logos, no case studies, no usage numbers. [site] [yc]
Price Personal: free (own keys). Company: $100 a month flat for the whole team, unlimited agent-hours, first month free, no per-seat price. Enterprise: contact sales (SSO, SCIM, self-host, VPC). [site]
Path Placeholder company name "Sean Grove NewCo" in the YC list [yc]. Chat product listed May 2026, macOS beta 20 May 2026, "Continuous Context Compiler" (C3) announced 15 Jun 2026 as coming soon (directory dates, not independently confirmed) [press]. The site now says "Now in beta" and C3 is still "coming soon" [site]. [yc] [press] [site]

2. Their proposition, reverse engineered

Part What they do
Who buys An engineering team that already runs several coding agents. The one who signs is the team lead or founder, at a $100 a month price (inference from the flat price) [site].
The pain Work is cheap now. The slow part is human decisions and checking. People jump between PR tabs, miss approvals in Slack and leave agents idle on small questions (paraphrase) [yc].
The promise Run dozens of agents from chat, from a laptop or a phone, and ship together. "The work of a team twice the size", not measured [site].
How it works Install a CLI and a macOS app. A local "Commander" starts Codex in folders you trust. The chat runs in the browser, the work runs on your machine. Agents get a small MCP tool set to read the thread and message the owner [docs]. Git snapshots save unsaved work before each agent turn [docs].
The output A chat thread per piece of work, with diffs, commands run, screenshots, test output and a replay, plus a "decision inbox" of calls that need a human [site] [yc].
Proof style None yet. Examples only, such as a 25-minute run that stalled at minute 3 [site]. The founder states a goal of "thousands" of agents by year end, not a result [yc].
Who sells it Self-serve sign-up and download, with contact-sales for Enterprise. Open jobs: 0 [site] [yc].
Price model Flat team price, free single-user tier. Users bring their own model keys [site]. A free-daily-model offer with Wafer.ai is a launch promotion [yc].
The story Enemy: the human as the bottleneck, and agent work hidden in terminals. Belief: a company should be steered like a model, using "deliberative alignment" ideas (from model training) applied to teams [yc]. Mission line: be the "CEO of their own 1,000-agent company" [yc].

3. Positioning

  • Category they claim: team chat for coding agents; "codex commander" [site]. Later: a "Continuous Context Compiler" [site].
  • Category a YC partner would put them in: developer tools, a control surface for coding agents (inference). Closest known peers named by one directory: Omnara, Limeline, Shipable [press].
  • What they want to be known for: the place where a whole team steers all its coding agents together, and where decisions become the source of truth that drives the next work.
  • Where they sit on the map: layers L1/L2/L3 = 1/2/1, buyer = sme (inference from price), developer.

4. Three layers vs ours

Layer Them (score + evidence) Us (from the system inventory, "Marton's logs show")
L1 Company brain 1. A "decision inbox" and threads agents can read over MCP are live [yc] [docs]. The compiled source of truth that "generates new work and verifies completed work" is on the site as "coming soon" [site] [yc]. Today it is chat history, not a lasting memory other work reads. A vault of about 1,400 pages, 112 call transcripts and about 8,700 customer quotes, searchable by meaning. The Partner (one agent across Macs) is only partly built.
L2 Agent orchestration 2. Many parallel agents, steered by a team, with human approval for network, credential and repo writes [site]. Only Codex is started by the product; Claude Code is "coming soon" [yc] [docs]. Routing is by people, not by a router. About 5,600 jobs (Codex 4,756, Claude 839, Gemini 46) across 3 Macs, with queue, claim, review and land steps. Routing between makers is by rules.
L3 Execution and learning 1. Agents do the work and the team checks it in the thread: diffs, test output, replay [site]. No sign of a measured result changing the next run. Checking is by humans. Every failure gets a record and an automatic check (about 160). Offer feedback lines fell from 44 to 18 as rules were added. Only about 15% of jobs reached the main code, and landing was unreliable in October.

5. Same and different

Topic Same or different Them Us
Core belief Same The bottleneck is human decisions and checking, not the typing [yc]. Same (brief: agents make a lot of work, teams can trust little of it).
Buyer Same Small engineering team already running agents [site]. W1: small startup team wanting 24/7 coding (brief). No customer yet.
Where work runs Same On the user's own machine, with the user's own keys [site] [docs]. Marton's logs show it runs on his own Macs, with his own accounts.
Who steers Different People steer in a chat, live, from a phone [site]. Marton's logs show a queue runs jobs while he sleeps and he reviews after.
How trust is built Different A human reads the diff in the thread [site]. Marton's logs show a different AI reviews each job, then tests gate the merge.
Learning Different Roadmap: decisions compile into specs [site]. Nothing shown yet. Marton's logs show failures turn into automatic checks. Learning loops are thin (1 to 4 weekly runs).
Makers supported Different Codex today, Claude Code soon [yc]. Codex, Claude and Gemini routed by rule.
Delivery Different Self-serve software, flat price [site]. A system run by two founders, nothing sold yet.
Proof Different No numbers [yc]. Real counts of jobs and merges, but the merge rate is low and the system is Mac-bound.
Size of ambition Different "1,000-agent company" for any team [yc]. One wedge at a time, SME buyer.

6. Wedge overlap

Wedge Fit Why
W1 Trusted 24/7 coding for small startup teams direct Same buyer and problem: agents make work and the team must direct and trust it [yc]. They answer with live human steering. W1 is "trusted while you sleep", checked by machines.
W2 Seeing whether agents produce adjacent Run logs, replay, "unblock the agents that need you" [site]. It shows what is stuck for runs a person is watching. No cost view, no silent-failure checks, no fleet report found (UNVERIFIED that none exists).
W3 Company memory from every call adjacent C3 is meant to join calls and transcribe them, answer with citations and reconcile conflicting plans [site]. For engineering teams, and not shipped.
W6 Tool for product managers adjacent "What to build, why it changed, who approved it" kept next to the work, and specs generated from it [yc] [site]. This is the PM's job inside an engineering chat. Coming soon, not shipped.
W8 The other 90% of an AI skill adjacent Decisions become a source of truth that "verifies completed work" [yc]. No claim about skill upkeep, feedback from use or retiring old skills (UNVERIFIED that none exists).

7. Whitespace (where they stop)

  • Layer gap: a human is the checker. Nothing shows machines checking work and learning from accepted and rejected runs [site]. To own it we need a run where a different AI plus tests catch real errors, and the next run is measurably better.
  • Buyer gap: they sell to teams that want to watch and steer. They skip the owner who wants it to run unattended and trusts only the result (inference from the live-steering pitch) [site]. To own it we need 2 to 3 small teams saying they will not watch agents.
  • Number gap: no throughput, merge rate or cost number anywhere [yc]. To own it we need one before/after number, such as merged changes per week or share of agent work that lands.
  • Price gap: not a gap. $100 a month for a whole team is cheap [site]. We can not undercut it, so price must come from the result.
  • Speed gap (not a gap): first value is fast: sign up, install the CLI, start a channel [docs]. We can not win on speed to first run. (Left out of the front matter on purpose.)

8. For the YC application

  • Proven today: Marton's logs show his system has already run about 5,600 agent jobs across three Macs and three AI makers, each job checked by a different AI before merge, with about 160 failure records that each added an automatic check. Linzumi's public material shows people steering and reading agent work, with no machine check or result number [site] [yc].
  • Draft answer to "how are you different from Linzumi?": DRAFT. "Linzumi lets a team watch and steer coding agents in a chat. We make the agents' work trustworthy without anyone watching: a different AI checks every job and every failure becomes a rule." Founders say it their own way.
  • Future claim (needs proof): "A small team goes to sleep and wakes up to merged, checked work, and the share that lands goes up each week." Needs one real team's repo run for 4+ weeks, with the share of agent work that merged before and after, and a GitHub-native runner (the inventory says the system is Mac-bound).
  • What we understand that they don't: the scarce thing is trust in finished work, not a better place to steer it. needs proof on a customer; Marton's logs show it on his own repo, where cross-checking and rules were built after repeated landing failures.
  • Copyability (M3): medium. C3 is a roadmap item that aims at checking and learning from decisions [site], so they want to go that way, and a funded team led by an alignment researcher could add machine review in months (inference). What slows them: the product is built around a person steering live from a phone [site], one maker (Codex) [docs], and users' own machines. A trusted unattended merge needs a loop of results data and a runner they do not have (inference). Not hard, so M3 can not score above 3 against Linzumi alone.
  • Why now (M4): None found. Their launch dates (macOS beta 20 May 2026, C3 announced 15 Jun 2026) are their own product moves, not a change that helps us more than them [press]. Do not use "agents can now do the work": that is general AI progress.
  • Do not claim: "we have more agents" or "better UI" (they have the UI and the funding). "Trusted 24/7 coding" as a result (our own merge rate is about 15% and landing failed for weeks in October). Anything about customers. "They only steer by hand" as fact (C3 roadmap says otherwise).

9. Ideas worth taking

  1. Free single-user tier and a flat price for the whole team. It removes the buying step for a small team [site].
  2. One thread per piece of work, holding the diff, the commands and the replay. A buyer can audit one thread without reading logs [site].
  3. A "decision inbox" that shows only calls a human must make. It turns noisy agent work into a short list [yc].

10. Unknowns

  • Real users, revenue or retention. Ask in the Launch YC thread or check their X account for posts about usage (public pages only).
  • Seed size and date. Only a directory gives Dec 2025, and amounts are not public. Look for a filing or announcement.
  • Whether C3 ships and what it checks. Watch linzumi.com and the demo videos named on the YC page.
  • Whether Claude Code support, a Windows or Linux app or a cloud runner is coming. The site shows macOS (Apple Silicon) only [site].
  • UK customers or staff. Nothing found.

Sources

Technical detail
Slug
linzumi
Lens hash
179096c479
Pass
deep
Stale
no