Mandri

Mandri is a local control plane for AI agents. It connects a desktop client and a command-line interface to a local daemon that manages sessions, providers and agent runtimes.

This is beta software. Interfaces may change between beta releases.

Installation

Install the Python package into an isolated tool environment with Python 3.13:

uv tool install --python 3.13 "mandri==0.1.0b0"

The package provides two commands:

mandri --help
mandri-daemon --help
mandri-daemon serve

The daemon stores configuration and its database in ~/.mandri. Use --base-dir to select a different data directory. Its default address is http://127.0.0.1:8787.

Provider setup

Register a provider and enter its API key at the interactive prompt:

mandri provider add openrouter
mandri provider list
mandri provider models openrouter

Credentials are verified when added. Use --key "$OPENROUTER_API_KEY" to supply a key from an environment variable, or --no-verify to skip the initial verification. To verify again later, run mandri provider verify openrouter.

The provider kind is inferred when its name matches a supported kind: openrouter, opencode, opencode_go, openai, anthropic, gemini, ollama, lm_studio or custom. For a different name, pass --kind:

mandri provider add my-provider --kind openrouter

For ollama, lm_studio and custom, also supply the API base URL with --base. All provider commands accept --base-dir to select a data directory other than ~/.mandri; use the same directory when launching a harness. Remove unused credentials with mandri provider remove <name>.

Run a harness

Choose a model ID from mandri provider models <name>, then launch a harness:

mandri run --model "openrouter/<model-id>" codex

Replace <model-id> with the exact model ID returned by the provider, including any slashes. The full reference is <provider-name>/<model-id>; for a provider registered as my-provider, use my-provider/<model-id>.

Supported harnesses are codex, claude, opencode, agy and pi. Codex uses the runtime bundled with Mandri. Claude Code and OpenCode must be installed and available on PATH; agy requires the Antigravity CLI and pi requires the Pi CLI.

mandri run starts the local daemon if needed and runs the harness in the foreground through the Mandri gateway. It uses the current working directory unless --cwd is set. Put Mandri options before the harness name, and pass harness-specific arguments after --:

mandri run --model "openrouter/<model-id>" --cwd /path/to/project codex
mandri run --model "openrouter/<model-id>" --effort high claude
mandri run --model "openrouter/<model-id>" codex -- --help

Reasoning effort support depends on the selected model. mandri run requires host execution with privacy mode set to none; if the configured protection requires managed execution, use mandri sessions start instead.

License

Mandri is licensed under Apache-2.0. See LICENSE for details.