Agent strategy for Backend Copilot: plans multi-step tool calls, executes them against your registered backend catalog, replans on errors. Read-only guardrails by default. Requires the backend-copilot tools plugin.
A Dify Agent strategy plugin: a plan → execute → replan multi-step pipeline that, together with the backend-copilot tools plugin, queries any registered backend API in natural language. Read-only guardrails by default.
This plugin is the strategy only — it does not include the API catalog or HTTP tools. Install and configure the backend-copilot tools plugin first (catalog YAML, credentials, and the tool all come from it). See its README.
Chinese docs: README.zh-CN.md
Source repository: https://github.com/zhou0928/backend-copilot
The strategy is driven by the pipeline modules under :
Install , then fill in the backend address, catalog YAML, and credentials on its authorization page. See the backend-copilot README.
Dify → Plugins → Install → Local file: . Or via GitHub: install from , pick the release.
Add an Agent node and pick the Backend Copilot strategy:
The strategy enables history-messages: in Chatflows, multi-turn conversation context is carried in.
Just ask in natural language. The Agent reads the catalog, plans, calls, and summarizes on its own — no per-API node wiring needed.
Reuse the three catalogs from the tools plugin's (paste into the "API catalog" parameter):
Q: "Backend tools not found"?
The Agent node's "Backend tools" parameter must have the tool from backend-copilot attached, and that plugin's credentials (backend URL + catalog YAML) must be configured.
Q: Planning keeps failing or steps are messy?
① Make sure the model supports tool-call (the strategy requires the scope); ② write clear for each API — plan quality depends directly on catalog descriptions; ③ use "Instruction" to constrain the output format.
Q: Budget exhausted mid-run?
Defaults are 600s / 20 steps. For complex queries, raise the execution budget and step cap, or narrow "Allowed tools" to reduce wasted attempts.
Q: Want to restrict the Agent to a few APIs?
Two ways: ① set an allowlist in "Allowed tools"; ② trim the catalog YAML to only the needed APIs (recommended — smaller catalog, better planning).
Q: Multi-turn context is lost?
The strategy enables history-messages; make sure conversation variables are wired into the Agent node in your workflow.
See LICENSE in the repository; privacy statement in PRIVACY.md.
Every outbound call builds its address at runtime, so no host can be listed.
Last checked 09/28/2026 01:45 AM · version 0.2.0