Donovan Agent

Bring the model that fits the work.

Donovan is provider-flexible rather than tied to a single hosted service. It supports OpenAI and OpenAI-compatible APIs, Anthropic, DeepSeek, Qwen, Ollama, LM Studio, and custom local or hosted /v1/chat/completions endpoints.

The setup wizard guides users through API keys, model selection, optional Tavily search, workspace permissions, and tool approval settings. Providers can be changed later without rebuilding the workspace.

donovan setup
donovan model
donovan model set

Context that survives beyond one prompt.

Long-running state is stored in SQLite, including sessions, message history, tool calls, audit logs, memories, learned skills, project context, scheduled tasks, and activity events. This allows Donovan to resume work, recall prior decisions, and keep a traceable record of actions.

Higher-level workflows turn raw tools into repeatable engineering processes. Contracts define a goal and success criteria; recipes capture reusable procedures; sandboxes stage commands before promotion; profiles preserve workspace preferences; and evaluations or agent tests check repeatable behaviour.

create a contract for fixing the browser companion setup
set up a recipe named release-check that runs tests and summarizes risks
configure router automatically for this project

Work across the terminal, browser, and external tools.

Donovan supports the Model Context Protocol for connecting servers that expose tools, resources, and prompts. MCP servers can be registered, inspected, diagnosed, and called from the interactive agent.

Browser work has two modes. Browser Companion works with already-open Chrome, Edge, Brave, Vivaldi, Opera, Arc, Chromium, and Firefox tabs through a local extension. Dedicated automation sessions use optional Playwright support for controlled browsing, text extraction, screenshots, and interaction.

Safari requires a separately packaged and signed Safari Web Extension, so it is not supported by the unpacked Browser Companion.

Capability stays visible and permissioned.

Because Donovan can edit files, execute code, and connect external tools, its defaults are designed around explicit workspaces and auditable actions. Workspaces must be approved, destructive commands are risk-classified, writes and patches are logged, tool results are recorded, API keys are masked in configuration output, and sensitive system paths are blocked.

Full autonomy should only be used on machines and folders where broad tool execution is acceptable.

A modular Python system built for extension.

The repository separates the agent loop from execution, providers, security, memory, browser tooling, MCP, planning, scheduling, skills, subagents, and terminal UI. This keeps integrations replaceable and makes security-sensitive behaviour easier to inspect.

Running donovan doctor checks Python, the operating system, terminal encoding, configuration, writable data directories, SQLite, provider connectivity, Tavily, the shell backend, Git, ripgrep, Node, workspace permissions, browser automation, checkpoints, skills, and registered tools.