Grounded investigations
Search a fixed local context index, process Alertmanager events, and keep operational evidence attached to a durable Run.
Open-source product by SysRoot
LLM-assisted operations with host-side control
Akritas connects an OpenAI-compatible model to selected local context, authorized tools, and a validated change pipeline without giving the model direct access to production or workspaces.
- Self-hosted
- Policy-controlled
- Evidence-first
requestInvestigate high CPU alert
contextlocal.runbooks.searchauthorized
toolmetrics.query_rangeread only
evidence3 host results attachedrecorded
runready for review
The model reasons and proposes. The host controls capabilities.
Akritas treats model output and external data as untrusted. The model never becomes the authority for host-side actions.
The trust boundary
The host selects context, authorizes tools, validates proposed changes, performs writes, and records results according to local policy.
Operational assistance you can inspect
Akritas is designed for bounded investigations and controlled file changes - not autonomous production access.
Policy-controlled tools
Expose only explicitly authorized MCP tools. The host validates schemas, permissions, arguments, and timeouts before execution.
Verifiable changes
Prepare exact file replacements in isolation, run validators, review a host-generated diff, and require one-time approval before writing.
Evidence and audit
Inspect what the host actually executed through Run records and a bounded append-only audit trail, rather than trusting model claims.
One boundary from request to evidence
Every capability remains behind the Akritas host throughout the Run lifecycle. Each step is bounded by local configuration and policy.
- 01Request
A person or Alertmanager submits bounded operational context.
- 02Reason
The model analyzes only the context and tool schemas supplied by Akritas.
- 03Authorize
The host applies local policy before retrieval, tool calls, validation, or writes.
- 04Record
Results and execution evidence are attached to a Run for review.
What Akritas does not hand to the model
The host keeps access, execution, approval, and audit authority outside the model boundary.
No direct production or workspace access
No arbitrary shell execution
No unreviewed file writes
No approval or audit authority
Start with the trust boundary, then run the quickstart
The documentation covers installation, model backends, MCP policy, local context, safe changes, Runs, and current limitations.
Open Akritas documentation