Overview
Leviath is a structured runtime for AI agents that perform multi-stage work. Instead of giving one model and one context window responsibility for an entire run, it defines the model, toolset and context budget for each stage in a single blueprint.
The runtime is distributed as a 74 MB Rust binary and is designed to host many agents in one process. Its approach separates stable task information from conversation and tool-call history, allowing each region to follow its own retention rules.
Key Features
- Context regions keep the task, plan and codebase notes pinned while conversation history compacts independently and tool calls move through a sliding window.
- Stages can select different models, tools and budgets, reserving more capable models for work such as implementation or deep review.
- A durable run journal records calls when they start and finish, so completed operations can be replayed after a restart instead of repeated.
- Blueprint graphs route progress, errors, retries, reassessment and human checkpoints according to the shape of the workflow.
- Fan-out stages let agents cover multiple files, topics or analysis areas concurrently.
Included Agents
Leviath ships with seven pre-built agents whose blueprints also serve as working examples. The catalog covers codebase changes, review, research, data analysis and log analysis, with model preferences that can fall back to any configured provider.
The coder agent discovers the repository, prepares a plan for approval, optionally prototypes uncertain approaches, implements the change and reviews the result. The reviewer uses a fast scan before deeper correctness, security and architecture analysis, then produces a ranked report.
API and Operation
The daemon can expose REST and WebSocket interfaces through its serving command, with a machine-readable OpenAPI contract kept aligned by tests. Requests require bearer-token authentication, while cross-origin browser access is disabled unless explicitly configured.
The server binds to the local interface by default and supports configurable request concurrency and deadlines. Administrators can optionally mount mutating routes, confine agent working directories, restrict unattended execution settings and disable blueprint seed commands for remotely started runs.