Model configuration

Mango's orchestration worker sends inference requests to your configured model endpoint. The current adapter requires a Messages-shaped POST /v1/messages API with streaming. An OpenAI-shaped endpoint alone is not sufficient.

Complete the Quickstart first. You need a reachable endpoint, a credential it accepts, and a model ID it serves. Real model calls may incur charges.

Create a local configuration

From the repository root:

make dev-env-init

Open ~/.config/mango/dev.env in your editor and set:

MANGO_API_KEY=sk-mango-local-development
MANGO_MODEL_BASE_URL=https://api.example.com
MANGO_MODEL_API_KEY=replace-me
MANGO_MODEL_ID=your-model-id
MANGO_MODEL_AUTH=x-api-key

Use the endpoint's base URL without /v1/messages. MANGO_MODEL_AUTH accepts x-api-key or authorization-bearer, depending on your endpoint.

The file uses literal NAME=VALUE lines, without shell quotes or export. It lives outside the repository and must have no group or other permissions. make dev-env-init creates it with mode 600 and leaves an existing file intact.

MANGO_API_KEY authenticates applications to Mango. MANGO_MODEL_API_KEY authenticates the orchestration worker to the model endpoint. Keep the latter out of client applications and sandbox containers.

Apply the configuration

make local-up
make local-health

make local-up loads the configuration file and rebuilds/recreates the local services as needed. This differs from the explicit offline command in the Quickstart. The API remains at http://localhost:8080.

Set the Agent's model to an ID served by this endpoint. The quickstart programs use offline-fake; use a real-model example below for your first model-backed Session. Existing Sessions retain their Agent snapshot, so updating an Agent alone does not change a Session you already created.

Try an application

Choose an example and follow its setup and run commands:

• Human-in-the-loop gate: let the model request an application action or human decision, then continue from the result.
• Specialist team: delegate a review and follow up with an existing specialist.
• Multi-agent SDK guide: use the same resource hierarchy from Go, Python, or TypeScript.

The example wrapper loads your local configuration:

scripts/with-dev-env make demo-hitl-gate

This command starts an interactive client, not the Mango services. The wrapper can use a different file through MANGO_ENV_FILE. See each example for its client runtime requirements and cleanup.

Enable tools

A text-only Agent needs no Environment worker. For bash, read, write, edit, glob, or grep in the default self_hosted Environment, start a Docker worker.

Web Search and Web Fetch execute at a supporting model endpoint and require always_allow. Remote MCP tools execute through Mango's orchestration runtime. Neither is redirected to the Docker worker. See where tools run.

For self-hosted Sessions, stage ordinary files and repositories in the operator-owned worker workspace. Memory Stores and pinned Skills are prepared through the worker protocol.

Troubleshooting

Keep orchestration workers on the same Temporal Task Queue consistently configured. For process configuration, see Deployment. Contributor-only endpoint and integration checks are listed in CONTRIBUTING.md.