fusionkit
Tools

Supported tools

The coding agents FusionKit can launch, and the behavior they share.

FusionKit launches an unmodified coding agent and backs it with a fused model panel. The agent speaks its normal wire protocol to a local gateway and never learns that fusion happened. Four agents share the neutral launcher path:

ToolCommandGateway dialectInstall
Codexfusionkit codexOpenAI Responses (/v1/responses)openai/codex
Claude Codefusionkit claudeAnthropic Messages (/v1/messages)Claude Code docs
Cursorfusionkit cursorOpenAI Chat, via the bundled Cursorkit bridgecursor.com/cli
OpenCodefusionkit opencodeOpenAI Chat through the generic driveropencode.ai

You only need the agent you plan to launch. fusionkit doctor checks which binaries are on your PATH before a run.

What every launcher shares

Each launcher composes the configured RouteKit router, Python synthesizer, and Fusion Harness Gateway, then receives the same neutral ToolLaunchSpec. One Ctrl+C tears down Fusion-owned processes. External RouteKit daemons remain.

FusionKit's own flags come first; everything after the first forwarded argument (or after --) goes to the agent untouched:

fusionkit claude --repo /path              # --repo is a FusionKit flag
fusionkit codex -- "fix the build"         # forwarded to codex

The behavior below works the same way in all four tools:

BehaviorHow
Fused modelThe agent sees fusion-panel as a normal model.
Ensemble modelsEach ensemble is fusion-<name>; the default is fusion-panel.
Rate-limit handoffA throttled passthrough turn reruns on the panel by default. See Rate-limit handoff.
Durable sessions--continue and --resume <id> restore a prior session.
Cost controls--budget <usd> caps gateway-observed spend. See Cost and models.
Observability--observe boots the local trace dashboard.
Sub-agentsEach named ensemble becomes a delegable sub-agent unless you pass --no-subagents.
Routing boundaryEnsemble members are live namespaced RouteKit model IDs. Use RouteKit for single-model launches.

The per-tool pages cover what differs: how each agent is wired to the gateway, which environment the launcher sets, how the model picker is populated, and the tool-specific caveats.

Other clients

Anything that speaks the OpenAI Chat Completions API can use fusion without a launcher. Run fusionkit serve and point the client at the printed URL with model fusion-panel. See Inference endpoint.

To add a launcher for a new tool, see Adding a new tool.