Configuring a runtime¶

A runtime is whatever LLM (or CLI agent) actually generates the replies in a chat. Hive doesn't talk to a single hard-coded provider — you configure as many runtimes as you want, and chats pick one per session.

Open Settings → Models to add or edit.

Settings → Models & runtimes tab

Supported providers¶

Provider	What it is	Auth	Tools?
`ollama`	Local Ollama daemon	none	text-only today
`openai`	OpenAI's API (or any OpenAI-compatible endpoint)	`OPENAI_API_KEY` env / pasted key	yes (MCP tools)
`anthropic`	Anthropic's API	`ANTHROPIC_API_KEY` env / pasted key	yes (MCP tools)
`openrouter`	OpenRouter aggregator	`OPENROUTER_API_KEY` env / pasted key	yes (MCP tools)
`claude-code`	Local `claude` CLI (uses your Claude Pro/Max sub)	depends on Claude CLI	yes (Claude Code's own tools)
`aider`	Local `aider` CLI	depends on aider's config	yes (aider's own tools)
`pi`	Local `pi` CLI	depends on pi's config	yes (pi's own tools)
`hive-daemon`	Remote `hived` instance	bearer token	passthrough
`custom`	OpenAI-compatible HTTP endpoint	varies	varies

claude-code, aider, and pi are the subprocess runtimes — Hive launches a local CLI. There is no generic subprocess provider; to point Hive at some other OpenAI-compatible server, use custom.

Provider presets¶

When you add a runtime in Settings → Models, the picker offers presets that pre-fill the endpoint and capability flags so you don't have to remember them:

OpenAI — api.openai.com.
OpenRouter — the OpenRouter aggregator.
Azure OpenAI — an Azure deployment endpoint (your https://<resource>.openai.azure.com URL + deployment name).
Ollama — your local Ollama daemon.
Custom — any other OpenAI-compatible HTTP endpoint; fill in the base URL yourself.

All five are OpenAI-wire-compatible, so the same multi-turn tool loop works across them. Pick the closest preset, then adjust the endpoint, model id, and API key as needed.

Runtime config schema¶

Each runtime is a [[runtimes]] block in hive.config.toml. The Settings UI is a GUI over the same fields:

[[runtimes]]
id = "anthropic-claude"
name = "Anthropic Claude"
provider = "anthropic"
kind = "remote"
endpoint = "https://api.anthropic.com"
preferred_model = "claude-sonnet-4-5"
supports_tools = true
performance_score = 9.5
cost_per_1m_input_tokens_usd = 3.0

For subprocess agents (aider, pi), endpoint is the binary path:

[[runtimes]]
id = "aider"
name = "Aider"
provider = "aider"
kind = "local"
endpoint = "/usr/local/bin/aider"
preferred_model = "gpt-4o"
supports_tools = true

See the Configuration reference for every field.

API providers — environment variables¶

For API providers, Hive can read the credential from an environment variable you control. Set it in your shell config before launching:

# ~/.zshrc
export ANTHROPIC_API_KEY="sk-ant-..."
export OPENAI_API_KEY="sk-..."
export OPENROUTER_API_KEY="sk-or-..."

Environment variables are a convenient fallback — quitting and relaunching picks up new ones. Note that a key you paste into Settings → Models is saved to disk in settings.json under the app data dir (so it persists across launches); it is stored in plain text, protected only by your OS account and disk encryption. Use the env-var path if you'd rather the key never be written by Hive.

Subprocess agents — installing the binary¶

Aider and pi are user-supplied CLIs; install them yourself and point Hive at the binary:

Aiderpi

pip install aider-chat
which aider
# → /usr/local/bin/aider

Then point Hive at /usr/local/bin/aider.

Install the pi CLI (an OpenAI-compatible coding-agent backend), then point Hive at its binary:

which pi
# → /usr/local/bin/pi

Hive bootstraps a temporary models.json and PI_CODING_AGENT_DIR so pi can reach your configured OpenAI-compatible endpoint (see crates/hive-runtime/src/provider/subprocess.rs).

For any other OpenAI-compatible server, don't install a binary — add a Custom runtime (provider = "custom") and fill in the base URL, as described under Provider presets.

Picking which runtime a chat uses¶

When you create a new chat, Hive picks the workspace's default_runtime. Switch per chat from the Primary runtime select just above the composer — it lists every configured runtime; choosing one re-points that chat.

To have a named agent always use a specific runtime, configure a Workspace Agent in Settings → Models and address it by its @handle from the composer.