03
A real agent, in a chat you own
Token-streamed answers, a live activity trace of every tool call, rich colored markdown and Mermaid, and per-turn cost — running on your infrastructure, against your data.
A neutron star doesn't outshine the field — it out-binds it. neutron core collapses a whole governed agent platform into one small, dense core, then holds N configured agents in orbit with gravity: roles, approval gates, and rules that never let power escape.
It isn't public yet — we'll send one note the moment the door opens. No spam, ever.
Everything a neutron star does with mass, neutron does with configuration.
A star collapses into a core a few kilometres wide — matter at its densest.
The whole platform is a thin chrome around the Claude Agent SDK — a core small enough to read, dense enough to run everything.
Immense gravity holds the star together. Nothing escapes it.
Approval gates, roles, and non-escalation rules hold every agent. No tool call, no config change, no privilege moves without them.
One spinning core sweeps many beams across space.
One image, deployed once, presents as many branded agents — each with its own persona, tools, channels, and repos. All of it is data.
A neutron star is layered — each shell a different state of matter.
Three seams, independently swappable. Chat surfaces, the agent engine, and your capabilities plug in — they're never welded in.
Most agent stacks treat governance as a safety valve. neutron treats it as the point — the binding force that makes autonomous agents safe to hand real access.
Risky tools stop at a human card in the chat — the exact command spelled out, a live countdown, Approve & run / Deny. Survives refresh; replayed into the trace.
A gated tool can require a role to approve — and the requester can never approve their own action. Role approvers get a cross-thread inbox. No admin break-glass.
Agents can manage agents — create them, tool them, scope them — but can never grant more privilege than they hold. Every mutation rides the gate by default.
Read-only by default, approval-gated in the middle, read-write when you mean it. Per instance, per tool. Read-only always wins.
Token-streamed answers, a live activity trace of every tool call, rich colored markdown and Mermaid, and per-turn cost — running on your infrastructure, against your data.
Every instance ships a built-in core agent and any number of named ones — each with its own persona, tools, guardrails, model, and workspaces. All configured, none forked.
Curated templates or from scratch, edited across Persona / Permissions / Tools / Model / Guardrails / Workspaces. Pin an agent to a conversation; hide admin-only ones.
Claude Opus · Sonnet · Fable on the Claude SDK; GLM and Kimi through the same driver. The OpenCode engine adds GPT-5.5 and any OpenAI-compatible model — everything OpenCode supports — plus AWS Bedrock. An engine badge on each.
Ask an agent to "learn the assay modules and remember them" and it self-authors a markdown guide — long-term domain memory it maintains and re-reads across turns.
The assay toolbox mounts ~55 infra/SaaS modules behind one approval-gated run tool; bake a module's docs into an agent's prompt, or declare CLIs by catalog pin or url+sha256.
The web UI is one surface. The same governed agent — same memory, same gates — reaches into your repos, your chat channels, and any MCP client.
@mention the bot in a GitLab or GitHub comment and it investigates and replies inline. Give it a set of repos and it works across them in its own git worktree — gated pushes and MRs, default-branch pushes always refused. Or plug the whole instance into any MCP client with a bearer token.
/mcp endpoint with an audited chat toolneutron core isn't public yet. Leave your email and we'll bring you into the private beta as the door opens — one image, one database, one config file away from your own governed agent.