fruxon-ai/fruxon
Build, test, deploy and improve AI agents on Fruxon from Claude: draft an agent in plain language, connect tools like Slack, GitHub and Jira, run evaluations and sandbox tests, publish revisions, trigger agents on schedules or events, manage knowledge bases, and debug any run from its execution trace. Requires a Fruxon account and the fruxon CLI.
The fruxon CLI's machine contract: what its exit codes (2, 10-16) mean, the {"error":{...}} envelope on stderr, commands that refuse without --yes (exit 16), and the JSON / NDJSON shapes that agents draft run, agents sandbox stream, agents sandbox test, and agents tests watch emit. Load when a fruxon command fails or refuses, when you need to parse one of those streams, or when scripting the CLI from CI.
Author a working Fruxon agent revision end-to-end via the CLI: discover, draft, test, create, deploy. Load when the user says "build a Fruxon agent", "create a revision", or starts authoring agent JSON.
Bootstrap a brand-new integration (with API tools or Python code tools) end-to-end via the CLI: schema discovery, body authoring, tools, then a clean handoff to the user for credentials. Load when the user asks to add an integration; the flow first checks whether one already exists and only builds a new one if nothing suitable is found.
Work out why a run did what it did, from its execution trace: find the run, read the step tree, drill into a tool call's parameters and result, and reconstruct the state the agent decided from. Load when a run failed, hung, or produced a surprising result — including the common case of a run that reports COMPLETED but did the wrong thing.
Orientation to Fruxon — what an agent / revision / definition / integration is, and the fruxon CLI's main entry points. Load when starting any task that involves authoring or operating Fruxon agents.
Wire integration tools into an agent: discover what's connected, attach the right tool, point it at a credential slot. Load when adding tools (Slack, GitHub, Jira, Drive…) to a Fruxon agent.