Supported agents & inference engines

Hive supports agent CLIs, inference engines, and OpenAI-compatible model gateways by backend ID. These are software integrations; inclusion does not imply endorsement or a partnership unless marked Partner. Support tiers follow Hive’s backend acceptance bar: T1 runs unattended in headless pods, T2 has Hive-wired local confinement or deny-listing, and T3 is experimental / opt-in for unconfined local use.

Agent CLI backends by support tier

T1 core — headless pod path

These backends are wired for unattended headless operation on the pod path.

Backend	Integration	Hive config	Notes
Claude Code	Anthropic agent CLI	backend: claude	Core CLI path with Hive prompt, permissions, audit, queue controls, and model recording.
Claude Code via LiteLLM	Claude Code pointed at a LiteLLM proxy	backend: litellm	Uses ANTHROPIC_BASE_URL / Hive LiteLLM settings; inherits Claude Code’s local confinement posture.
GitHub Copilot CLI	GitHub agent CLI	backend: copilot	Hive manages repository scope, permissions, and audit trails for Copilot-driven work.
OpenAI Codex CLI	OpenAI agent CLI	backend: codex	Passes Hive-selected model and reasoning-effort settings where supported.
Goose	Block / Agentic AI Foundation agent CLI	backend: goose	T1 on the pod path; local mode remains experimental because Hive cannot wire a local OS sandbox for Goose.

T2 supported — confined or deny-listed local path

These backends have a Hive-wired local confinement floor: an OS sandbox or an enforced deny-list for host-state commands.

Backend	Integration	Hive config	Notes
Claude Code / LiteLLM	Anthropic CLI, optionally through LiteLLM	backend: claude / litellm	Uses Claude Code’s native sandbox locally; LiteLLM keeps the same posture.
OpenAI Codex CLI	OpenAI agent CLI	backend: codex	Hive narrows local workspace access with Codex sandbox flags.
GitHub Copilot CLI	GitHub agent CLI	backend: copilot	Hive uses Copilot sandboxing and explicit repository scope.
Muse Code	Meta agent CLI	backend: muse	Supported when installed by the operator and launched with Hive’s local controls.
OpenCode	OpenCode terminal coding agent	backend: opencode	Hive wires a host-state command deny-list rather than claiming OS confinement.

T3 experimental — explicit opt-in for unconfined local use

These backends are available but are experimental where Hive cannot wire local confinement. Local launch requires each backend’s explicit HIVE_<BACKEND>_DANGEROUSLY_RUN_UNCONFINED=1 opt-in; container mode remains preferred.

Backend	Integration	Hive config	Notes
Goose (local)	Block / Agentic AI Foundation agent CLI	backend: goose	Experimental on contributor-local mode; T1 on the pod path.
Google Antigravity CLI	Google agent CLI	backend: agy	Hive manages unattended mode and reasoning effort; local confinement is not wired.
IBM Bob	IBM watsonx Code Assistant CLI / bobshell	backend: bob	Separate from the watsonx.ai inference gateway; headless use requires API-key auth.
Pi	pi-coding-agent CLI	backend: pi	Long-running interactive backend with Hive kicks when ready.
Aider	Aider CLI	backend: aider	Uses provider credentials and Hive-provided repository context.
Kilo Code	Kilo Code CLI	backend: kilo	Experimental local path; containerized operation is preferred.
Oh My Pi	Oh My Pi CLI	backend: omp	Hub-agent method and experimental contributor-local backend.
OpenHands	OpenHands CLI, headless-only	backend: openhands	New / experimental (hive#10634). Contributor relay, headless mode; T3 refusal-gated locally (HIVE_OPENHANDS_DANGEROUSLY_RUN_UNCONFINED=1) because the bare CLI has no sandbox Hive can wire. Not in the contributor image or the K8s headless allowlist.

Go-side backend

Backend	Integration	Hive config	Notes
Gemini CLI	Google Gemini CLI	backend: gemini	Go-side manager launch today; it has no contributor-relay wiring, so local-path support tiers do not apply until that is added.

Inference gateways and engines

Gateway backends are not agent binaries. Hive routes Claude-style agent calls through an OpenAI-compatible translation path to these endpoints, so CLI support tiers do not apply.

Gateway	Hive config	What Hive does
vLLM	backend: vllm	Routes inference to a vLLM OpenAI-compatible endpoint for self-hosted models.
llm-d	backend: llm-d	Routes inference to distributed Kubernetes model serving.
LiteLLM	backend: litellm / gateway route	Centralizes model routing and keys behind a LiteLLM endpoint.
IBM watsonx.ai	backend: watsonx	Uses watsonx.ai’s OpenAI-compatible gateway and IBM project / token settings.
OpenRouter	backend: openrouter	Uses OpenRouter as a named gateway with operator-selected model IDs.
Custom OpenAI-compatible	kind: custom route	Lets operators point Hive at any compatible /v1/chat/completions endpoint.

Model providers and classifiers

Integration	Hive config	Notes
Jev	classifier.backend: jev	TypeSafe AI — Partner. Optional v6 advisory smart classifier.
Anthropic	backend: anthropic	Provider surfaced through the configured CLI or gateway.
OpenAI	backend: openai	Provider surfaced through Codex, gateway routes, or compatible proxies.
DeepSeek	backend: deepseek	Provider surfaced through a configured OpenAI-compatible gateway such as LiteLLM.

TypeSafe AI is a Hive Commons partner.

Planning, work sources, and related ecosystem

Hive sits in a broader agentic-maintenance ecosystem. These projects are either integrated today, tracked as related work, or useful context for operators comparing orchestration approaches.

Infrastructure thanks

Hive Commons thanks Akamai (Linode) and Oracle Cloud (OKE) for Kubernetes infrastructure donated through the CNCF, Cloudflare for DNS, edge and tunnels, GitHub Copilot for AI inference donated to CNCF projects through the CNCF, and Bluehost for domain hosting.

Source control, work sources, and sign-in

The product landing page groups runtime integrations separately from the agent CLI and inference list above. Current shipped surfaces are:

• Source control: GitHub and GitHub Enterprise are the production GitHub App path. The forge-neutral adapter layer includes GitLab and Gitea / Forgejo issue and change-request flows as those paths graduate.
• Work sources: GitHub Issues (default), GitHub Projects, Linear, Jira Cloud, Jira Data Center / Server (deployment: datacenter), Spektacular run stages, and Crustify / Wavefront migration graphs.
• Sign-in: Microsoft, GitHub, IBM, Google, and other OIDC providers by hub configuration.

Jira Data Center / Server support uses Jira REST API v2, preserves context-path base URLs, and supports Personal Access Token bearer auth or basic auth fallback. Custom CA and TLS trust settings are still in progress and are not listed as shipped stable functionality.

Recent Hive platform additions

• Swarm mode: per-repository swarms with idle-unlock, themes, Discord announcements, and /api/leaderboard/swarm.
• The Commons: contributors can subscribe to multiple hives, rank them, and choose ranked, spread, or neediest routing; hivectl hives web opens a local management UI.
• Teams and achievements: distro / OS / agent team leaderboards are exposed at /api/leaderboard/teams, and Achievements 2.0 adds Solo, Dual, Fireteam, and Raid tiers plus the local-model track.
• Edge-only v6: admin MCP (/api/admin/mcp, cmd/hive-admin-mcp) and the dashboard Extensions tab for Spektacular are on the v6 edge branch until that line is promoted.