Skip to content
OmniConflux

Your machines. Your models. Your call.

The OmniConflux Confluence mark: seven glass strands, each carrying its own light, braided into one ring.

Your agents. Rival checks.

OmniConflux is designed to coordinate your agents on your machines, have a different vendor check each task, and leave you to decide what ships.

Pre-register for the closed betaSee a recorded run

Closed beta. Only your work email and your goal are required.

Any model. Any harness. Your login, never ours.

Plug in what you already run. OmniConflux is designed to run any command-line agent and bring models in three ways, so joining a fleet does not mean replacing your setup.

Command line

Any CLI agent, run as a process.

MCP

Designed for deeper control: state, handoff, approvals.

ACP

The same deeper control.

One fleet socket.

Your login.The vendor’s own tool, on your machine, as the vendor allows.

Your API key.Direct to the vendor.

Your local model.Ollama or llama.cpp on hardware you own.

No lossless context across vendors. Agents will share work through git worktrees and files, and some need a short setup step.

Pre-register for the closed beta See what connects

Your machines are the fleet.

OmniConflux is designed to pool the hardware you already own for agent work: each paired machine reports what it can offer, and a coordinator hands each task to the machine that fits. Your tower is not a paperweight.

Sample fleet. Sample data, not a live node list.

One goalMigrate billing to the new API.

  1. Your laptopCLI agent, your login
  2. Your towerMCP agent
  3. Your serverCLI agent, API key
  4. Your GPULocal model

Each machine will dial out to the relay, so a box behind a home router needs no open inbound port.

Pre-register for the closed beta See how a fleet fits together Build a sample fleet

This is a real review from this site's own build. In the design, OmniConflux gives each finished task the same check: a verifier from a different vendor returns APPROVE or REWORK with reasons. An agent that grades its own work tends to agree with itself, so do not let it.

A rival checks every diff.

Recorded run P1a

BuilderClaude familyWrites the change

  1. glossary
  2. Removed: built natively into OmniConflux surfaces

    Gemini family: presents Aura as shipped. status lists Aura as Not Live. Say "planned" or "designed to".

  3. Added: A planned... network..., designed around on-device topic classification. Not live yet.
  4. use cases computer use voice
  5. Removed: compiles... into a multi-step execution graph across your fleet

    Gemini family: The real voice tool sends a command to a single harness.

  6. Removed: Every cursor movement... streamed securely to your phone

    Gemini family: Keep "every action shown", which the task asked for, but drop the streaming...

  7. Claude family: I removed the execution graph, accessibility tree and DOM inspection, phone streaming, remote abort, sandbox and secondary-display claims, the...

VerifierGemini familyDifferent vendor

Round 1 ran

  • $ npm ci
  • $ npx tsc
  • $ npm run build
  • $ node scripts/check-copy.mjs
  • $ node scripts/check-ssr.mjs

REWORK

19 defects, each with a reason.
The change goes back.

Round 2 ran the same checks

APPROVE

Reworked. Checked again.
No defects.

19 defects back to the builder

  1. 1The draft goes out for review.
  2. 2A verifier from another vendor runs the checks.
  3. 3REWORK: 19 defects, each with a reason.
  4. 4The builder rewrites the flagged lines.
  5. 5The same checks again. APPROVE.

Quoted from the run file; paths, hosts and file names are redacted, and earlier rounds were not retained. Playback compressed. All product names are trademarks of their owners.

Verification is a check with recorded evidence, not a promise that every result is correct.

Pre-register for the closed betaSee how a rival check works

You stay in control.

In the design, the last word is yours: not the agent’s, and not the rival’s.

Pre-register for the closed beta See what is live and what is not
Sample approval prompt. Phone approval is planned, not shipped.

A rival checked the pages of this site and sent them back.

A builder from one vendor wrote seven pages. A verifier from another vendor returned 19 defects, most of them claims nothing backs. After the rework, it approved. Press play.

Recorded run: final two rounds of a three-round review.

Note from the run file: earlier rounds were not retained. Timing is compressed; paths, hosts and file names are redacted.

Pre-register for the closed betaRecorded run data (JSON)Ledger source (JSON)
Recorded runAPPROVED

Lane P1a

machine 1

  1. 00:00

    goal: replace the StubPage stubs with real server rendered content for: use cases, use cases computer use voice, glossary, press, status, enterprise, earn network governance. Read first: <file> (including the v2 addendum and "AJ decisions 2026-10-01"), <file> (surface specs) and the existing page <file> for component...

  2. 00:01

    verification starts, retained round 1 [agent C: Gemini family]

  3. 00:02

    $ npm ci [agent C: Gemini family]

  4. 00:03

    $ npx tsc [agent C: Gemini family]

  5. 00:04

    $ npm run build [agent C: Gemini family]

  6. 00:05

    $ node scripts/check-copy.mjs [agent C: Gemini family]

  7. 00:06

    $ node scripts/check-ssr.mjs [agent C: Gemini family]

  8. 00:07

    verdict: REWORK, 19 defects [agent C: Gemini family]

    1. use cases computer use voice: Line 19: the intent "compiles... into a multi-step execution graph across your fleet". The real voice tool sends a command to a single harness.
    2. use cases computer use voice: Line 26: "queries the operating system accessibility tree and browser DOM structure... within an isolated sandbox window". Nothing implements this.
    3. use cases computer use voice: Line 33: "Every cursor movement... streamed securely to your phone". Keep "every action shown", which the task asked for, but drop the streaming...
    4. use cases computer use voice: Line 40: the verifier "confirms... no unintended state changed, and tests reproduce reliably". This overclaims; the sibling page says verification...
    5. use cases computer use voice: Line 53: "file deletion or external network requests require explicit confirmation" is stated as a fixed behaviour with no source.
    6. use cases computer use voice: Lines 56-58: Remote Abort Controls have no source.
    7. use cases computer use voice: Lines 61-63 and 104: sandbox boundaries, "dedicated secondary displays, virtual desktop spaces, or headless sandboxes" have no source.
    8. use cases computer use voice: Line 71: "prioritizes native CLI and DOM manipulation" has no source.
    9. use cases: Line 29: "deterministic reproducer scripts" as a guaranteed outcome.
    10. use cases: Line 30: "AST diff inspection".
    11. use cases: Line 55: "strictly bounded operating system sandboxes".
    12. use cases: Line 74 (FAQ): the verifier "performs AST analysis, and confirms deterministic behavior".
    13. use cases: Lines 84 and 112-113: the intro says "every step" is verified (every agent task is the backed claim), and the FAQ says critical file changes and network requests...
    14. enterprise: Line 112: the verifier is "executing your internal CI test suites and AST checks". Same AST claim as above.
    15. enterprise: Lines 31-36: the Managed Cloud Fleet is described in present tense ("provisioned, monitored, and scaled", "dedicated sandboxing per worker node", "GPU acceleration...
    16. glossary: "built natively into OmniConflux surfaces" presents Aura as shipped. status lists Aura as Not Live. Say "planned" or "designed to".
    17. status: "agents operate entirely on-device... does not depend on cloud API uptime for core orchestration or verification" is a product behaviour claim about software that has...
    18. earn network governance: Line 35: "Granular category blocklists" is not in the Aura <file>, which lists off by default, an inspector and one toggle.
    19. earn network governance: Advertiser vetting is thin: only "vetted developer tool affiliate programs" at line 42. The task asked how vetting works. Add one or two plain lines...
  9. 00:08

    builder starts on machine 1 [agent B: Claude family]

  10. 00:09

    rework starts [agent B: Claude family]

  11. 00:10

    $ npm ci [agent B: Claude family]

  12. 00:11

    $ npx tsc [agent B: Claude family]

  13. 00:12

    $ npm run build [agent B: Claude family]

  14. 00:13

    $ node scripts/check-copy.mjs [agent B: Claude family]

  15. 00:14

    $ node scripts/check-ssr.mjs [agent B: Claude family]

  16. 00:15

    use cases computer use voice: I removed the execution graph, accessibility tree and DOM inspection, phone streaming, remote abort, sandbox and secondary-display claims, the... [agent B: Claude family]

  17. 00:16

    use cases: removed AST inspection, "deterministic reproducer scripts", "strictly bounded operating system sandboxes", and the approval prompts for file and network actions. The... [agent B: Claude family]

  18. 00:17

    enterprise: Removed AST from the FAQ and from the audit-log roadmap item. [agent B: Claude family]

  19. 00:18

    glossary: Aura is now "A planned... network..., designed around on-device topic classification. Not live yet." The contextual ad broker entry is reworded the same way. [agent B: Claude family]

  20. 00:19

    status: the cloud-outage FAQ now says "designed so that..." and notes the client software is not released. [agent B: Claude family]

  21. 00:20

    earn network governance: The whole page now uses "designed", "planned" or "will". [agent B: Claude family]

  22. 00:21

    verification starts, retained round 2 [agent C: Gemini family]

  23. 00:22

    $ npx tsc [agent C: Gemini family]

  24. 00:23

    $ npm run build [agent C: Gemini family]

  25. 00:24

    $ node scripts/check-copy.mjs [agent C: Gemini family]

  26. 00:25

    $ node scripts/check-ssr.mjs [agent C: Gemini family]

  27. 00:26

    verdict: APPROVE [agent C: Gemini family]

  28. 00:27

    lane P1a approved

00:28

Swarm ledger

Lanes that built this site, counted from the surviving review logs.

22 lanes. 23 verification rounds.

19 APPROVE and 4 REWORK verdicts in the surviving logs.

Counted separately: 6 REWORK verdicts found in embedded prompt reports.

Six defects the verifiers caught

  1. P1aAura described as shipped

    glossary: "built natively into OmniConflux surfaces" presents Aura as shipped. status lists Aura as Not Live. Say "planned" or "designed to".

  2. P1aUnsupported browser and system access

    use cases computer use voice: Line 26: "queries the operating system accessibility tree and browser DOM structure... within an isolated sandbox window". Nothing implements this.

  3. L4Branch did not merge or compile

    The branch doesn't merge cleanly or compile against current main. L1 merged its server component (ba69560, 23:21) 14 minutes before L4's last commit (23:35), and L4 never...

  4. L4Live mark would not start

    After the merge, the live mark would never start on any page.

  5. L4Ring rendered as an overexposed blur

    The live ring renders as an overexposed blur. The strands can't be told apart. This happens on SwiftShader and on a real GTX 1060, while the reference renders a crisp braid on...

  6. I1Decorative icons repeated alt text

    No icon is marked decorative, so screen readers repeat the same long alt text.

How the ledger is counted

Counts are lower bounds where logs were overwritten. Surviving verify files and embedded prompt reports are counted separately. Git counts cover only main..swarm lane ranges; merged commits are absent. Ordinary commits do not establish review rounds. Do not add these overlapping sources together.

Full transcript, 28 events

RECORDED RUN: final two rounds of a three-round review

  1. 00:00

    goal: replace the StubPage stubs with real server rendered content for: use cases, use cases computer use voice, glossary, press, status, enterprise, earn network governance. Read first: <file> (including the v2 addendum and "AJ decisions 2026-10-01"), <file> (surface specs) and the existing page <file> for component...

  2. 00:01

    verification starts, retained round 1 [agent C: Gemini family]

  3. 00:02

    $ npm ci [agent C: Gemini family]

  4. 00:03

    $ npx tsc [agent C: Gemini family]

  5. 00:04

    $ npm run build [agent C: Gemini family]

  6. 00:05

    $ node scripts/check-copy.mjs [agent C: Gemini family]

  7. 00:06

    $ node scripts/check-ssr.mjs [agent C: Gemini family]

  8. 00:07

    verdict: REWORK, 19 defects [agent C: Gemini family]

    1. use cases computer use voice: Line 19: the intent "compiles... into a multi-step execution graph across your fleet". The real voice tool sends a command to a single harness.
    2. use cases computer use voice: Line 26: "queries the operating system accessibility tree and browser DOM structure... within an isolated sandbox window". Nothing implements this.
    3. use cases computer use voice: Line 33: "Every cursor movement... streamed securely to your phone". Keep "every action shown", which the task asked for, but drop the streaming...
    4. use cases computer use voice: Line 40: the verifier "confirms... no unintended state changed, and tests reproduce reliably". This overclaims; the sibling page says verification...
    5. use cases computer use voice: Line 53: "file deletion or external network requests require explicit confirmation" is stated as a fixed behaviour with no source.
    6. use cases computer use voice: Lines 56-58: Remote Abort Controls have no source.
    7. use cases computer use voice: Lines 61-63 and 104: sandbox boundaries, "dedicated secondary displays, virtual desktop spaces, or headless sandboxes" have no source.
    8. use cases computer use voice: Line 71: "prioritizes native CLI and DOM manipulation" has no source.
    9. use cases: Line 29: "deterministic reproducer scripts" as a guaranteed outcome.
    10. use cases: Line 30: "AST diff inspection".
    11. use cases: Line 55: "strictly bounded operating system sandboxes".
    12. use cases: Line 74 (FAQ): the verifier "performs AST analysis, and confirms deterministic behavior".
    13. use cases: Lines 84 and 112-113: the intro says "every step" is verified (every agent task is the backed claim), and the FAQ says critical file changes and network requests...
    14. enterprise: Line 112: the verifier is "executing your internal CI test suites and AST checks". Same AST claim as above.
    15. enterprise: Lines 31-36: the Managed Cloud Fleet is described in present tense ("provisioned, monitored, and scaled", "dedicated sandboxing per worker node", "GPU acceleration...
    16. glossary: "built natively into OmniConflux surfaces" presents Aura as shipped. status lists Aura as Not Live. Say "planned" or "designed to".
    17. status: "agents operate entirely on-device... does not depend on cloud API uptime for core orchestration or verification" is a product behaviour claim about software that has...
    18. earn network governance: Line 35: "Granular category blocklists" is not in the Aura <file>, which lists off by default, an inspector and one toggle.
    19. earn network governance: Advertiser vetting is thin: only "vetted developer tool affiliate programs" at line 42. The task asked how vetting works. Add one or two plain lines...
  9. 00:08

    builder starts on machine 1 [agent B: Claude family]

  10. 00:09

    rework starts [agent B: Claude family]

  11. 00:10

    $ npm ci [agent B: Claude family]

  12. 00:11

    $ npx tsc [agent B: Claude family]

  13. 00:12

    $ npm run build [agent B: Claude family]

  14. 00:13

    $ node scripts/check-copy.mjs [agent B: Claude family]

  15. 00:14

    $ node scripts/check-ssr.mjs [agent B: Claude family]

  16. 00:15

    use cases computer use voice: I removed the execution graph, accessibility tree and DOM inspection, phone streaming, remote abort, sandbox and secondary-display claims, the... [agent B: Claude family]

  17. 00:16

    use cases: removed AST inspection, "deterministic reproducer scripts", "strictly bounded operating system sandboxes", and the approval prompts for file and network actions. The... [agent B: Claude family]

  18. 00:17

    enterprise: Removed AST from the FAQ and from the audit-log roadmap item. [agent B: Claude family]

  19. 00:18

    glossary: Aura is now "A planned... network..., designed around on-device topic classification. Not live yet." The contextual ad broker entry is reworded the same way. [agent B: Claude family]

  20. 00:19

    status: the cloud-outage FAQ now says "designed so that..." and notes the client software is not released. [agent B: Claude family]

  21. 00:20

    earn network governance: The whole page now uses "designed", "planned" or "will". [agent B: Claude family]

  22. 00:21

    verification starts, retained round 2 [agent C: Gemini family]

  23. 00:22

    $ npx tsc [agent C: Gemini family]

  24. 00:23

    $ npm run build [agent C: Gemini family]

  25. 00:24

    $ node scripts/check-copy.mjs [agent C: Gemini family]

  26. 00:25

    $ node scripts/check-ssr.mjs [agent C: Gemini family]

  27. 00:26

    verdict: APPROVE [agent C: Gemini family]

  28. 00:27

    lane P1a approved

Build your fleet.

Pick machines and a model for each, send a sample goal, and watch a rival from another vendor return REWORK, then APPROVE. No signup.

SAMPLE RUN. Sample data. No real agents run. Nothing you choose here is sent unless you claim a fleet seat and save it.

Pick your machines.

Choose the machines in your sample fleet. Any combination works.

Your sample fleet: Laptop, Desktop tower. Ready to receive a goal.

Pick your models.

Choose a model vendor on each machine. Mix vendors in one fleet.

Laptop
Desktop tower
Give your fleet a goal.

This is a sample. The agents are not real.

Sample runReady when you are.

  1. Press Run sample to watch your fleet work.

    1. The goal fans out to every machine that runs an agent.
    2. A rival from a different vendor checks the first machine's work.
    3. It sends the work back with a reason, or approves it.
    4. You approve last.

SAMPLE DATA. Not a real run.

Pre-register without the sample

Sample only. In the design, finished work comes back for your approval. Phone approval is planned. Nothing happened here.

Vendor names are trademarks of their owners. Naming a vendor describes designed compatibility, not endorsement.

No landlord.

Your keys.

Your models.

Your work is designed to land as commits in your own git repositories and files on your own disks, so you can swap a model, change a harness or leave without moving anything. Your work was never ours to keep.

Read the sovereignty principles

Pre-register for the closed beta.

Start with an email and a goal. Everything after that is optional.

Use the address where you want access updates.

What do you want your agents to work on? (required)

Choose any that fit, write your own goal, or do both.

We store your email, your goal, your marketing choice if you tick it, the time and form source, and HMAC hashes of your IP address and browser user-agent instead of the raw values. We use them to plan the closed beta, email you about access and prevent abuse. Delete it any time: email us. Read our privacy information.

Questions before you register

What is OmniConflux?

An agent platform designed around machines you own, models you choose, independent verification, on-device context and voice. It is in closed beta, and the client software is not released yet.

How is agent work checked?

The design sends each task to a verifier from a different vendor, which returns approve or rework with reasons. It is a check with recorded evidence, not a guarantee that every result is correct.

What does private by default mean?

The design keeps code, files and local-model work on machines you control. When you choose a cloud model, that vendor receives what the agent sends it, directly from your machine, under its own terms.

Does OmniConflux lock me into one AI vendor?

Not by design. The design brings models in three ways: the vendor’s own tool on your machine under your login, your own API key, or local models on your hardware.

Can I use the AI subscription I already have?

On your own machines: OmniConflux is designed to run the vendor’s own installed tool under your login, as the vendor allows. By design, it does not pool personal accounts or pass a seat login through a hosted backend.

Which agents does it run?

Any command-line agent is designed to run at the process level. Agents that speak MCP or ACP are designed to get deeper control over state, handoff and approvals. In the design, agents share work through git worktrees and files, and some agents need a short setup step.

Can you run the fleet for me?

Self-serve is the default design. A managed fleet is planned, not live, and is a talk-to-us option. It will run only on your API keys or enterprise tokens, never on personal subscription seats.

Can I download OmniConflux today?

No. OmniConflux is in closed beta and the client software is not released. Pre-registration is required. Share your email and what you want your agents to work on, and we will email you about access as the beta develops.

How much will it cost?

Free to start, plans announced at launch.

How does earning work?

Aura by OmniConflux is a planned ad network and is not live. Revenue share is announced at launch.