Mezmo open-sourced Aura, an incident agent defined entirely in TOML

Mezmo, the observability company, released Aura on GitHub under Apache 2.0. It is a Rust-based agent platform for investigating and remediating production incidents, and the main claim on the repository is that the whole agent system is defined in a single reviewable TOML file. The repository has gained 267 stars, and the Hacker News launch thread was posted this morning.

What it is

Aura correlates data from the sources an SRE already relies on during an incident: logs, traces, metrics, and deployment history. Given a fresh alert, an Aura agent goes through those data sources and produces a recommendation, which can include actions like rolling back a recent deployment. Integrations cover observability tools, cloud platforms, and infrastructure systems. Model support is broad: OpenAI, Anthropic, Bedrock, Gemini, Ollama, and OpenRouter are named on the readme, so a team can use Aura with whichever provider its policy allows, including a local Ollama instance for environments with no network connection.

The TOML config, and why it matters

The distinguishing design choice is that the agent, its tools, its data sources, and its escalation paths are all declared in TOML. That means the agent's behaviour goes through the same review process as any other config in the repository: a pull request changes it, a code review approves it, and a commit history explains why it changed. Two other pieces make this workable in production. Sensitive actions need human approval, so an agent can propose a rollback but a person still authorises it. Every step the agent takes is exported as an OpenTelemetry trace, so the same observability stack that watches the application also watches the agent that reasons about it.

The audit log is the feature

For a team that has been curious about agents in the incident-response path but worried about an agent taking actions in production, the combination of a text-based agent definition, an approval step on the actions that matter, and traces of every decision is the design that makes a trial possible. The problem with earlier agents of this kind has been that the reasoning is hidden, the actions cannot be undone, and the config is kept in an interface no one reviews. Aura moves each of those into a form the team already handles for the rest of its systems.

The other useful point is that Mezmo, a company whose business is observability, chose to publish the agent core as open source rather than as a hosted feature. That does not by itself say the code is well engineered, and 267 stars is early. It does say the choice of running the agent inside a customer's own infrastructure was intentional, and the release under Apache 2.0 lets a team fork the agent definition without committing to Mezmo's platform.

Source

Primary source: mezmo/aura on GitHub. Community discussion: Hacker News.