neutron core
The Neutron Core agent engine runs governed work across tools, repositories, schedules, and peer instances. Configure many agents from one core image and choose the adapter each one uses.
A governed turn, end to end
Every entry path resolves the acting identity and grants before work begins, then keeps the decision and outcome attached to the same activity trace.
-
01
Request
Work arrives from the web product, a private message, a GitHub or GitLab mention, a token-authenticated client, or a durable schedule.
-
02
Identity and grants
Neutron resolves the named agent, acting user, provider, capabilities, workspace resources, and safety posture for that turn.
-
03
Tool decision
The Neutron Core agent engine resolves the adapter first. Claude Agent SDK adapter actions meet an explicit allow, gate, or deny rule; the OpenCode SDK adapter's different native enforcement path remains a visible boundary.
-
04
Approval when required
A gated action pauses with the exact operation and routes to the configured role or named users. Ungated work does not invent an approval step.
-
05
Trace and outcome
The persisted decision, actor, timeout, usage, cost, and resulting message or run remain connected in the activity trace.
Governed autonomy is the product
The Neutron Core agent engine applies controls at the point of action through its Claude Agent SDK adapter. Neutron-owned connectors use shared policy checks; native OpenCode SDK adapter tool-governance parity remains in progress.
Three safety modes
readonly, approval-gated, and
readwrite set the default posture for each agent.
Tool-level decisions
Allow, gate, or deny individual tools. New MCP capability tools begin gated until an operator changes policy.
Routing when required
Route a gated action to a role or named users. Two-person separation is available when the policy calls for it.
OpenCode enforcement differs
The OpenCode SDK adapter currently does not share the Claude Agent SDK adapter's per-tool approval enforcement. Treat that difference as an explicit trust boundary.
Proactive work, without notification noise
Neutron can run headless turns on a clock, retain their results, and decide how urgently those results should reach an operator.
Durable schedules meet an interrupt governor
Five-field cron schedules start durable, headless agent turns. Retries survive a pod restart, results stay in per-schedule threads, and typed inbox cards preserve the outcome.
- The interrupt governor applies quiet hours and priority breakthrough before raising a toast
- A Constellation control core can delegate chat or administration to registered peers with scoped tokens
- Operators can gate schedule mutations and peer administration with per-tool policy
- Headless and repository-channel turns deny gated actions when no live approval bridge is available
Capabilities and providers are controlled resources
Operators register integrations once, scope credentials and grants, then equip only the agents that need them.
MCP servers, tools, and skills
Register OAuth-aware MCP servers, pinned CLI tools, and reusable skills. MCP and skills support both adapters; CLI grants currently use the Claude Agent SDK adapter.
Credentials without redeploys
Scope and rotate named API-key credentials for compatible backends. Manage subscription and catalog connections through the OpenCode SDK adapter.
Secrets stay governed
Use OpenBao or the assay-engine encrypted vault. Sensitive capability values remain secret references instead of raw registry data.
One event contract
The Neutron Core agent engine normalizes output and usage from its Claude Agent SDK and OpenCode SDK adapters behind one event contract. Their native tool-governance paths are not yet equivalent.
Repository work stays isolated
GitHub and GitLab workspaces resolve through agent-scoped connections and persistent per-agent worktrees.
One agent identity per working context
Each agent receives only its granted repositories, guides, connections, and capabilities. A shared bare cache keeps checkout efficient while each working tree remains isolated.
- Investigate, branch, commit, and push according to workspace autonomy and the selected adapter
- PR/MR creation is governed separately and requires host CLI authentication
- Claude Agent SDK adapter workspace guardrails refuse default-branch pushes
- GitHub and GitLab mentions reply inline with conversation memory and cost attribution
Issue autopilot is under development
Today, Neutron responds to GitHub and GitLab mentions and can work inside assigned workspaces. Durable issue claiming, reconciliation, job state, and automatic branch-to-PR lifecycle are roadmap work; the product does not present arbitrary backlog pickup as shipped.
Available now
Mention-driven repository investigation, configured tool use, isolated workspace changes, and inline replies.
Roadmap
Label-driven issue intake, named-agent dispatch, durable progress updates, CI reconciliation, and operator-controlled merge handoff.
Operate agents on your terms.
Request access to evaluate Neutron against your infrastructure, identity, provider, and repository requirements.