XMarts room

Fall 2026

herdr

building the open agent runtime

Visit website

What they do

open runtime that keeps coding agents alive and shows which are blocked or done

How close to us

4, very close

Attention queue of working, blocked and done agents is our W2 problem; developer buyer, no cost or output measure.

Wedges

  • W2 Seeing whether your AI agents actually produce direct
  • W1 Trusted 24/7 coding for small startup teams adjacent
L1 Company brain (knows the organisation) 0, none
L2 Agent team (runs agents, tools, approvals) 2, clear
L3 Does and learns (finishes, checks, measures, updates the brain) 0, none

Who buys

Firm size
sme
Buyer type
developer
Role
individual developer who installs it; engineering lead if it spreads (inference)
UK
not known
Their own description (from YC)

Developers don't run one coding agent anymore. They run fleets of them — for hours, sometimes days — across laptops, servers, and sandboxes. Keeping track of that work is the new bottleneck: which agent is blocked, which one finished, which terminal is it in, and why did everything die when the laptop closed.

Herdr is the open-source home for that work. A background server owns the agents' terminal sessions on each machine, so the work survives disconnects, closed lids, and dropped SSH. Herdr detects each agent and turns the chaos into an attention queue — working, blocked, done, across every project — so you look only when an agent needs you. Every client is just a view: the bundled terminal UI, SSH from another machine, your phone. Close one, open another; the work never stops.

Software can drive it too: a CLI, a socket API, and a plugin marketplace that passed 500 community plugins in its first month. The runtime is Apache-2.0 and stays open. Four months after launch: 25k GitHub stars, 340k downloads.

Full teardown

herdr (YC Fall 2026): teardown vs XMarts

In one line: a free, open-source background server that keeps coding agents' terminals alive and shows which are working, blocked or done, for developers who run many agents. Threat: 2 because it owns the free "which agent needs me" screen for developers with huge reach, but it stops at terminal state: no output, no cost, no team view.

1. Fact file

Item Finding Source
What they say "Run them anywhere. Leave them running." YC line: "building the open agent runtime" [site], [yc]
Founded, team, funding Founded 2026, Ankara, Turkey. YC lists team size 1. They claim a $6M seed led by Bessemer, with YC and e2vc, plus angels (Tobi Lutke, Dane Knecht, Gorkem Yurtseven), announced 8 Sep 2026. The seed post says a "small team" is being built and hiring is now closed. One aggregator in search results listed $500k; not checked, treat as noise [yc], [blog:seed], [press:runtimewire]
Founders Can Celik, solo founder. YC says he built AI-assisted ultrasound technology and led AI teams in fintech. YC partner: Pete Koomen. The fact that likely made YC say yes: a solo project that went to 25k stars in about four months (inference, from their own traction claim) [yc], [blog:yc]
Customers None named. No logos or case studies. Users are individual developers. Traction they claim: 44,627 GitHub stars, 1,644,092 installs, 1,789 community plugins, 22 agent CLIs detected (site, v0.9.5). Earlier claim on 6 Aug 2026: about 25k stars, 340k downloads, 500+ plugins [site], [github], [blog:yc]
Price Core is free. Herdr Cloud is "coming soon" with a waitlist and no price. README gives a hey@ address for enterprise or partnership enquiries. One review said a commercial licence exists by email; that review says AGPL, which conflicts with the GitHub page (Apache-2.0) [site:cloud], [github], [press:bitdoze]
Path Early 2026: tmux-style agent multiplexer, built in Rust (a source says created March 2026; UNVERIFIED). 27 May 2026: live updates without killing processes. 10 Jun 2026: "Coding agents are becoming runtimes" thesis post. At some point the licence moved from AGPL to Apache-2.0 (their post says "switch from AGPL", no date). 3 Aug 2026: 95% less CPU. 6 Aug 2026: joins YC. 7 Sep 2026: v0.9, many machines in one client. 8 Sep 2026: seed [blog:index], [blog:yc], [blog:cloud], [blog:seed]

2. Their proposition, reverse engineered

Part What they do
Who buys Nobody pays yet. The user is a developer who runs several coding agents (Claude Code, Codex, Cursor, opencode and others) on a laptop, a server or a sandbox [site], [yc]
The pain "Which agent is blocked, which one finished, which terminal is it in, and why did everything die when the laptop closed." [yc]
The promise Agents keep running when you disconnect, and an attention queue tells you which one needs you [yc], [site]
How it works A Rust background server owns each agent's terminal. It reads each pane to mark working, blocked or idle. Optional per-agent hooks sharpen the signal. Clients (terminal UI, SSH, phone) are only views. A CLI and socket API let agents and software drive it, so agents can start, prompt and wait on each other [site], [yc], [press:bitdoze]
The output A live status list across projects and machines. No report, no number, no cost line (none found on site or in the review) [site], [press:bitdoze]
Proof style Public counts: GitHub stars, installs, plugins, agent CLIs supported [site]
Who sells it Nobody. Pure product-led, open source. Plugins are found through a GitHub topic with "No submission, no review queue" [site:plugins]
Price model Planned: free runtime, paid features for connecting many machines and clients (Cloud relay, end-to-end encrypted, "would not host your agents") [blog:yc], [blog:cloud]
The story Enemy: fragmentation and terminal-screen guessing. Belief: developers own their working environment, and the runtime should not be a private vendor detail; agents should compete on models and taste, not plumbing [blog:runtimes]

3. Positioning

  • Category they claim: "the agent runtime" and "the open agent runtime" [site], [yc].
  • Category a YC partner would put them in: open-source developer infrastructure, a tmux for coding agents, with a Docker-style bet that a hosted layer pays later (inference, from the free-runtime-plus-paid-connector plan [blog:yc]).
  • What they want to be known for: the free, neutral place where every coding agent lives.
  • Where they sit on the map: layers L1/L2/L3 = 0/2/0, buyer = sme, developer.

4. Three layers vs ours

Layer Them (score + evidence) Us (from the system inventory, "Marton's logs show")
L1 Company brain 0. No memory of the company or the work. Config and layouts persist, but no knowledge layer [site] Marton's logs show a vault of 1,396 pages, 112 call transcripts and 8,763 customer quotes, searched by meaning
L2 Agent orchestration 2. Hosts many agents from many makers, and an attention queue hands blocked ones to the human. Agents can prompt and wait on each other through the API, but herdr does not route or approve for them: the user or the agent does [site], [yc] (score kept at 2 by judgement; 1 is arguable) Marton's logs show 5,641 jobs run across three Macs and three engines (Codex 4,756, Claude 839, Gemini 46), each built in its own copy of the code and reviewed by a different AI
L3 Execution and learning 0. It does not do the work and does not check it. State detection reads the screen or hooks, and says nothing about whether the output was good [press:bitdoze] Marton's logs show automatic tests and about 370 checks before merge, and about 160 written failure records each tied to a check. Weak spot: only 15% of jobs reached main

5. Same and different

Topic Same or different Them Us
Problem Same slice Which agent is stuck or done [yc] W2 starts with "what is stuck", then adds "what silently failed" and "what it costs" (brief)
Belief Different The runtime must be open and neutral; plumbing is not where agents compete [blog:runtimes] A status light is not an answer; the question is whether the work landed (needs proof outside Marton's own use)
Buyer Different The individual developer, free [site] A small engineering team that wants agents shipping overnight (W1/W2 in the brief)
Unit of view Different A terminal pane and a machine [site] A job and its result (Marton's logs show 5,641 jobs tracked)
Owner of the result Different None. It keeps agents alive; the human owns the output We would own a number on a workflow (needs a customer run)
How they capture Different Screen reading plus optional hooks [press:bitdoze] Jobs go through our own queue, so we know the state without guessing (Marton's logs show)
Kind of work Same Coding agents [site] Coding agents, from three makers
Size Similar Solo founder, small team forming [blog:seed] Two founders, one user so far
Delivery Different Install script, Mac/Linux/Windows [site] Mac-bound, launchd, Marton's paths only. Nobody else has run it
Price Different Free, paid Cloud later, no price [site:cloud] Not set. The brief floats a price tied to coding hours gained minus token cost

6. Wedge overlap

Wedge Fit Why
W2 Seeing whether your AI agents actually produce direct Same user pain on "what is stuck". They show terminal state; they show no cost, no output check and no silent-failure list (none found) [site], [press:bitdoze]
W1 Trusted 24/7 coding for small startup teams adjacent Same agents, same developers. They keep agents running; they do not review or land the work. A possible layer beneath us or beside us (inference)

W8 (the other 90% of a skill) and W9 do not touch them. Their plugin marketplace is code plugins, not skill upkeep [site:plugins].

7. Whitespace (where they stop)

  • Layer gap: they stop at "alive and waiting". No checking of the output, no learning. To own this we must show a view of whether work landed, and a rule that changes after a rejected run (needs a customer run with a rework count falling).
  • Buyer gap: they serve the single developer and the multi-machine hobbyist. A team lead who needs a shared view across people and queues is not served yet (inference; Cloud may move there). To own it we need 3 to 5 small engineering teams saying the pain out loud.
  • Number gap: no cost, no output rate, no "done but not merged" count (none found). To own it we need one number on real agents for 4 weeks, such as spend per merged change. Marton's logs show 15% of our own jobs reached main, which shows why "done" is not a result, but it is one user.
  • Price gap: free is hard to beat for individuals. A paid, team-level view is the only price gap; their Cloud price is unknown.
  • Speed gap: they win here. Install is one command and works on day one [site]. We would have to match it, and today ours is Mac-bound.

8. For the YC application

  • Proven today: Marton's logs show a heartbeat that reports what is stuck, what silently stopped and what it costs, across 5,641 jobs on three Macs and three engines. Herdr's public pages show agent state (working, blocked, done) but no cost or output measure [site].
  • Draft answer to "how are you different from herdr?": DRAFT. "Herdr tells you an agent is blocked or done. Our system tracks whether the work actually landed and what it cost, and Marton's logs show that only a small share of agent jobs do." Founders say it their own way.
  • Future claim (needs proof): "Herdr is where agents run; we are how a team knows what they produced." Proof needed: one outside team's agents watched for 4 weeks, with a before/after number such as silent failures caught or spend per merged change.
  • What we understand that they don't: a "done" signal is not a result: Marton's logs show 42% of our jobs ended wont-merge and only 15% reached main (proven, but one user, our own system).
  • Copyability (M3): medium. Their plugin API and socket make a cost or output view quick to add (1,789 plugins in months) [site]. But output and "did it land" need the job and review data from a queue, which a terminal runtime does not own (inference). They are also pulled to stay neutral and light ("one rust binary, no electron") [github]. Evidence is partial, so not rated hard.
  • Why now (M4): None found.
  • Do not claim: that herdr has no cost or output tracking (we only found none on their site and in one review); that we beat their reach; that any team wants a shared agent view yet; that they are a rival to W1; that their funding is $500k or that their licence is AGPL (sources conflict).

M2 (what YC funds now) is NOT judged here. It comes from the themes of many companies in Pass 3.

9. Ideas worth taking

  1. A public support matrix that separates verified agent signals from screen guesses [blog:runtimes]. A W2 version could label every failure class "seen by hook" or "guessed".
  2. Zero-review plugin discovery through a GitHub topic [site:plugins]. A cheap way to get 1,000+ extensions. It also shows how far a free tier spreads.
  3. Free core, paid connector. Charge only for linking machines, not for running agents [blog:cloud]. A model for us: free local view, paid team view.

10. Unknowns

  • Whether Cloud will add team features, cost or output views, and at what price. Watch herdr.dev/cloud and the blog.
  • Real team size now (YC says 1; the seed post says hiring closed). Check the YC page again before submission.
  • Licence today (YC post says moved from AGPL to Apache-2.0; GitHub page shows Apache-2.0; one review says AGPL). Read the repo LICENSE file.
  • Stars: 25k (6 Aug) vs 44.8k (now). Install count 1.64M is their claim and not checked.
  • Any company paying or any team using it at work. None named.
  • Whether agent vendors are opening lifecycle hooks (their 10 Jun post argues for it). If a dated vendor change exists, it could be an M4 lead; not verified here.
  • Reception on forums (no Hacker News thread found in search).

Sources

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