Backend Copilot Agent

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

Connection requirements

This strategy plugin contacts only the backend registered in the companion tools plugin's catalog; it requires no separate credentials or endpoints of its own.
The model you select in the Agent node must support tool-calling; the tools plugin must be authorized with a reachable backend (use for host services from the Dify container).

How it works

The strategy is driven by the pipeline modules under :

Module	Responsibility
	LLM generates a structured JSON plan; each step declares which tool to call and with what params
	Runs the plan, sequential or parallel scheduling, trips the circuit breaker on budget exhaustion
	Re-plans from the scratchpad on step failure or invalid plans, with its own budget
	Intermediate results between steps; later steps can reference earlier outputs
	Time budget () and LLM call budget ()
	Filters available tools by allowlist
	Step-level executor abstraction, swappable implementations

Usage

1. Prerequisite: install & configure the backend-copilot tools plugin

Install , then fill in the backend address, catalog YAML, and credentials on its authorization page. See the backend-copilot README.

2. Install this plugin

Dify → Plugins → Install → Local file: . Or via GitHub: install from , pick the release.

3. Configure an Agent node in your workflow

Add an Agent node and pick the Backend Copilot strategy:

Parameter	Required	Default	Description
Model	✅	—	An LLM that supports tool-call
Query	✅	—	User question, e.g. "How many open tickets this week?"
API catalog (YAML)	✅	—	Paste the catalog YAML (same content as in the tool credentials); templating supported
Backend tools	✅	—	Attach the tool from backend-copilot
Read-only mode		on	On: APIs marked are always blocked; off: they may run only when the user explicitly asks
Allowed tools		empty	Allowlist; empty = all attached tools
Instruction		—	Answer style, constraints, e.g. "output as a table, amounts to 2 decimals"
Max plan steps		20	Maximum steps in one plan
Max replans		3	Maximum automatic replans
Execution budget (seconds)		600	Timeout circuit breaker
Output variable			Variable the final answer is written to, for downstream nodes

The strategy enables history-messages: in Chatflows, multi-turn conversation context is carried in.

4. Ask

Just ask in natural language. The Agent reads the catalog, plans, calls, and summarizes on its own — no per-API node wiring needed.

Read-only guardrails (three layers)

Catalog layer: APIs are read-only by default; write APIs must be explicitly marked in the catalog YAML
Credential layer: the tools plugin's "Allow write" master switch defaults to
Strategy layer: "Read-only mode" defaults to on, blocking all write APIs; when off, they run only on explicit user request

Example catalogs

Reuse the three catalogs from the tools plugin's (paste into the "API catalog" parameter):

— BladeX (tickets/workflow, 6 read-only APIs)
— RuoYi (users/departments/roles etc.)
— JSONPlaceholder (public demo, no auth)

FAQ

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.

Local development

License

See LICENSE in the repository; privacy statement in PRIVACY.md.