Context is lost between chats
Every session starts from zero. The codebase, the decisions and the conventions live in someone's head — usually the person who is now the bottleneck.
Living Intelligence
A living organisation of AI agents.
meLink neuron is the control room for a team of AI agents that plans, delegates, verifies and remembers — on your machine, on your models, with every move on screen.
Living Intelligence is a metaphor for good management, not a claim of sentience.
Engineering proof at a glance
62k
lines of TypeScript
145
test files
874 + 70
tests recorded, Windows + WSL
0.087 ms
median indexed read
636
run folders of dogfooding
127.0.0.1
binds locally by default
The bottleneck
You don't have an AI problem. You have a coordination problem.
One assistant can write a function. It cannot hold a plan, divide work, check its own evidence and pick up where it stopped. That is management work — and it is where AI projects stall.
Parallel work becomes manual work
You copy tasks between tools, reconcile edits and become the integration layer. The more agents you add, the more human glue you need.
“Done” is not evidence
A confident summary is not a passing test. Without fingerprints and exit codes, review is vibes with better typography.
The shift
From assistants to organisations.
Living Intelligence
An organisation of agents that plans, delegates, verifies, remembers and recovers.
2022–23
Prompt
2024–25
Workflow
2026 →
Organisation
| Dimension | Chat assistant | Agent workflow | meLink neuron |
|---|---|---|---|
| Unit of work | one reply | one pipeline | a team of agents |
| Who plans | the human | the graph | an orchestrator |
| Who holds state | the chat | the workflow run | Hubs, tickets and a knowledge base |
| Parallelism | none | fixed branches | dynamic, ownership-aware delegation |
| Verification | the human reads it | step success | fingerprinted evidence + exit codes |
| Oversight | read the answer | inspect the run | live graph, transcripts, artefacts, replay |
| Recovery | start again | re-run the job | pause, resume, replay, fail over |
Living Intelligence is a metaphor for good management, not a claim of sentience. Read the category thesis
What it is
A control room, not another chat window.
One goal enters. A plan, a team, a record of evidence and a replay come out.
Transcript
orchestrator → designer
“Own brand + layout for the pricing page. Declare the files you will touch.”
Verification
✓ typecheck exit 0
✓ 874 server exit 0
○ e2e gated —
Artifacts
- ①Every delegation is an edge you can click. Messages, handoffs and results are inspectable, not inferred.
- ②Verification is a list with exit codes. Not a summary paragraph written by the agent being reviewed.
- ③Replay scrubs the whole run. From 1× to 128×, with skip-idle, from the same event log the live view used.
Hub — the persistent organisation
Managers, projects, tickets, memory and history that survive restarts. One Main Orchestrator, one Project Manager per real folder, specialists on demand.
Use a Hub for ongoing delivery.
Runs — one-off and replayable
A single orchestrated job with recording and replay. Every run is persisted to SQLite and reconstructed from the same event log.
Use a Run for a single question.
How it works
Six moves, every one visible.
No new methodology. The same things you would ask of a good team.
Goal
Create a Hub for ongoing work or a Run for one job. Describe the outcome in plain language.
what you see · one goal
Plan
The Main Orchestrator writes the plan, splits it into tickets and sets dependencies.
what you see · ticket tree
Delegate
Managers spawn the roles a ticket needs, with owned paths and a definition of done.
what you see · 3 workers live
Verify
Declared paths are fingerprinted, raw exit codes recorded, stale evidence flagged.
what you see · exit 0
Remember
Results are written back to the project knowledge base and ticket memory.
what you see · KB updated
Recover
Providers fail over, usage limits pause work and storage pressure holds it. Nothing is silently lost.
what you see · run resumed
Capabilities
The parts that make it hold up.
Each is in the product today. Each proof line is a fact, not a promise.
Live agent graph
Every delegation, message, handoff and result is an edge you can click. Swimlanes and follow-activity keep a busy team readable.
HubGraph.tsx · SwarmGraph.tsx
Persistent Hubs
One Main Orchestrator, one Project Manager per real folder, tickets with dependencies, a tasks board and chat.
hub/hub.ts
Evidence-based verification
Declared paths are content-fingerprinted before tests run. Raw exit codes are recorded. Stale evidence is detected.
verificationPolicy.ts
Provider-agnostic by design
Ollama Cloud, local Ollama, OpenAI, Anthropic or any OpenAI-compatible endpoint. Provider sets fail over; limits pause work.
providers.ts · modelSets.ts
Project memory
A scout scans the repository into a project knowledge base; ticket memory and a global KB stop later work re-learning the codebase.
hub/kb.ts · roles/scout.md
Secrets and tool scoping
An AES-256-GCM vault per manager, scoped placeholders resolved only into URLs, and per-agent tool grants over a loopback MCP bridge.
hub/vault.ts · mcpBridge.ts
Operator channels
A Telegram console for one approved owner, cron jobs, Jira and Confluence over MCP, and per-project reverse-proxy publishing.
telegram/service.ts · crons.ts
Local by default
One Node process bound to 127.0.0.1, SQLite with WAL, no hosted account for state. Your machine is the server.
app.ts · storage/database.ts
Current boundaries: Telegram live delivery needs a real bot and account; Jira and Confluence are validated with mocks, not live servers. macOS is unverified. Some workflows need gh, Caddy, Docker CLI or Chrome/Edge.
Where it fits
Built for work that has to survive review.
The sharpest fit is software delivery, because the evidence is executable. The same primitives travel further.
Product engineering
The job
Ship a real repository with design, code, review and tests.
What neuron does
One manager per repo, specialist workers, ticket dependencies and verified acceptance.
Honest limit
Agents run with broad permissions and are not path-sandboxed; use version control.
Agencies & studios
The job
Run several client projects without context bleed.
What neuron does
Multiple Hubs, one manager per project folder, per-project knowledge bases.
Honest limit
Hubs share installation-wide concurrency, provider and usage budgets.
Platform / DevOps
The job
Automate changes that need an audit trail.
What neuron does
Replayable runs, recorded exit codes, git branch, commit and pull-request flow.
Honest limit
Needs gh, Caddy or Docker CLI for some workflows.
Solo technical founders
The job
Be the whole team without losing the thread.
What neuron does
The orchestrator plans, workers execute, transcripts and artefacts stay on disk.
Honest limit
Single-user by design; no multi-tenancy.
Regulated internal tooling
The job
Keep code and transcripts inside the network.
What neuron does
Local execution, local models and an encrypted secrets vault.
Honest limit
The SQLite database is not encrypted; treat the host and backups as sensitive.
Research & data
The job
Reproduce and inspect long-running work.
What neuron does
Replay from 1× to 128×, artefact capture and local transcription.
Honest limit
Some features need extra binaries or model downloads.
These are designed use cases, not customer case studies.
The record
Numbers from the repository, not the pitch deck.
Internal engineering evidence, dated and reproducible. These are not customer outcomes.
| Metric | Figure | Source in repository | Date |
|---|---|---|---|
| Application source | 61.8k lines TypeScript | server/src · web/src · shared/src | 2026-11-07 |
| Test suite | 145 test files, ~22.6k lines | repo manifests | 2026-11-07 |
| Recorded test run | 874 server + 70 web passing, 2 skipped | .melink/reports/cross-platform/validation.md | 2026-11-07 |
| SQLite read latency | 0.087 ms median / 0.130 ms p95 at 100k rows | .melink/reports/sqlite-verification.md | 2026-11-07 |
| Dogfooding footprint | 636 run folders, 18.2 GB, 3 Hubs | data/ | 2026-11-07 |
| Launcher start | 1.3 s cached start | .melink/reports/cross-platform/validation.md | 2026-11-07 |
Internal reports from development machines, Windows + WSL Linux, dated 2026-11-07. Not third-party benchmarks. macOS validation pending. Figures are a dated snapshot, not an SLA.
What we do not claim
×No hosted SaaS
×No SSO or per-user permissions
×No security certification
×No audited third-party benchmarks
Why now
The agents arrived. The control plane didn't.
Enterprise AI spend grew 5.4× in a year. Eight in ten tokens on OpenRouter now come from agents. Three in four builders use them — but only one in three uses more than one together.
Enterprise AI deployment spend
$177B
by September 2026, 5.4× the $33B a year earlier
Budgets are compounding. The constraint is execution quality, not willingness to buy.
Menlo Ventures, “2026: The State of Enterprise AI”, 3 Nov 2026.
Read the source
Agents as share of tokens
80%
of OpenRouter token usage; up 37× since February 2026
Agents are now the dominant workload, not an experiment beside the chat box.
Menlo Ventures, “2026: The State of Enterprise AI”, 3 Nov 2026.
Read the source
Builders using agents vs multiple agents
74%/ 32%
use agents to manage workflows; only 32% use multiple agents together
The 42-point coordination gap is the orchestration whitespace.
Menlo Ventures, “2026: The State of Enterprise AI”, 3 Nov 2026.
Read the source
The constraint moved from capability to coordination.
Read the researchPackaging
Software that runs where the work runs.
Indicative packaging · priced per deployment
Pay for a deployment, not a demo. How we intend to sell, not a published price list.
Design partner
IndicativeFor a small number of teams with a real project and a named owner.
- Private deployment with a weekly working session
- Direct influence on the roadmap
- A written evidence pack at the end
- Fixed annual fee, priced per deployment
- Limited places; no logo rights assumed
Private deployment licence
IndicativeFor one team or department that needs the control room on its own infrastructure.
- Annual licence plus seats and usage bands
- Provider setup and role configuration
- Knowledge-base onboarding and upgrade guidance
- Support terms agreed in the contract
Platform — later
RoadmapFor smaller teams that want to start without a deployment project.
- A self-serve or hosted option
- The same orchestration and verification semantics
- On the roadmap, not on sale
- Not pre-sold as if it exists
Pricing principles. We charge for a deployment and the people who use it, not per token. Model costs stay with the customer's provider account. No fake “unlimited”. No published price table until the unit economics are measured.
Every package: local-first, provider-agnostic, evidence-based verification, MCP bridge, no per-project cloud dependency.
The horizon
The living organisation.
Every team gets a control room. The intelligence is real. The life is a metaphor. The accountability is the product.
2026
Today
The control room: a team of agents you can watch, steer, verify and replay.
2027
Next Vision
Organisation memory across projects: shared role packs, cross-project knowledge, multi-machine orchestration — and a founding engineer hire.
2028
Horizon Vision
Governed autonomy: policy boundaries, approval gates and exportable audit trails.
The metaphor is a metaphor. neuron does not claim sentience or autonomy without oversight. The team at meLink is founder-led: we designed, built and dogfooded the product end to end. Full bios and references are shared during diligence.
Next step
Bring one real project. Watch it run.
A 30-minute private demo: your repo or a sample, screen-shared. You leave with a run recording and a cost estimate.
01
Pick the project. One folder, one person who owns the outcome.
02
Run the room. Create a Hub, bind a manager, brief the orchestrator, watch the team.
03
Inspect the evidence. Graph, transcript, tool call, verification record and replay.
Private deployments only. We never ask for production credentials. Prefer to read first? Download the PDF deck, one-page summary (PDF), or read the research.
Questions
The ones we get asked.
Is this a chatbot?
No. You give a goal to an orchestrator that writes a plan, opens tickets and delegates to specialist agents. Chat is the steering wheel, not the product.
Does my code leave my machine?
The app, the database and the agent engine run locally on 127.0.0.1. Model calls go to whichever provider you configure — Ollama Cloud, local Ollama, OpenAI, Anthropic or any OpenAI-compatible endpoint. You choose; local models keep inference local too.
Which models can I use?
Any provider above, subject to the provider's own terms and availability. Providers are grouped into sets with failover, and usage limits pause work instead of failing tasks.
How is “done” verified?
Declared source paths are content-fingerprinted before tests run, and raw exit codes are recorded. Stale evidence is detected. Full end-to-end runs are gated to one manager-nominated acceptance boundary, so “done” and “accepted” stay separate. It proves what ran, not that the tests were adequate.
How close is it to production?
It is a working private deployment used daily in our own delivery work: 62k lines of TypeScript, 145 test files, and 874 server plus 70 web tests recorded on Windows and WSL. Today's boundaries are explicit: one installation per team, no SSO or per-user permissions, a trust-based execution model, and a SQLite database that is not encrypted. Those are the next milestones on the roadmap, not claims we make now. Use it on trusted projects with version control.
Does it work on macOS?
Windows and WSL Linux are verified. Native macOS validation is on the roadmap; until it ships, treat macOS as unverified.
Is it open source?
No. It is a proprietary product. Access is through a private deployment or design partnership; the packages are pre-1.0.
What runs in the cloud?
Only the model calls you configure. Transcripts, tickets, knowledge bases, artefacts and the SQLite database stay on the machine. The database itself is not encrypted — treat backups as sensitive. Secrets in the vault are AES-256-GCM encrypted and never sent to models.
Can it run on our infrastructure?
Yes — that is the private-deployment model. One Node process, one SQLite file, your provider keys, your network. Some features need external binaries (gh, Caddy, Docker CLI, Chrome/Edge).
What does it cost?
Pricing is set per private deployment: an annual licence plus seats and usage bands. Design-partner pilots are a fixed fee, also priced per deployment. We do not publish a price list while the unit economics are being measured; a design partnership is the fastest route to a real number.