# pi-cursor-cli-provider

A [Pi Coding Agent](https://github.com/badlogic/pi-mono) custom provider that routes model requests through the **Cursor CLI**, enabling you to use your Cursor subscription inside Pi. This project is a heavily modified fork of [netandreus/pi-cursor-provider](https://github.com/netandreus/pi-cursor-provider).

# Motivation (if you don't feel like reading all this, go straight to [Installation](#installation)) I spent a long time looking for a proper Cursor provider for Pi, but the ones that worked best were too "hacky" in my opinion - they used the reverse-engineered API to call Cursor endpoints directly, masquerading as the legitimate Cursor CLI. They also had various issues, such as not passing the system prompt properly. I also tried to write my own extension based on the [ACP protocol](https://agentcommunicationprotocol.dev/). I even got it to the point where I could use it with Cursor - but Cursor does not expose all models through ACP. And the ones it does expose are available in only one variant - for example, only with "medium" thinking. That makes this approach, in my opinion, unusable. After all that, I stumbled upon [netandreus/pi-cursor-provider](https://github.com/netandreus/pi-cursor-provider). It worked, and most importantly, it used the legitimate Cursor CLI without any hacks. But it was not super polished - tool call rendering was basically just raw text, all session history was sent as a prompt on every turn, images were not supported, etc. Still, the idea was cool, so I decided to build on top of it. # Installation ### npm ```bash pi install npm:@akepka/pi-cursor-cli-provider ``` ### git ```bash pi install git:github.com/Strus/pi-cursor-cli-provider ``` For the extension to work, **you need to have [Cursor CLI](https://cursor.com/docs/cli/installation) installed and authenticated**. Authentication can be done either with the `agent login` command or with a token generated from the Cursor dashboard and set as the `CURSOR_API_KEY` environment variable. See the [Cursor CLI docs](https://cursor.com/docs/cli/reference/authentication) for more information. # Usage Just start Pi and everything should work. Models are auto-discovered on startup, so you can choose them with `/models` or `/scoped-models` as usual. Pi's thinking effort settings are mapped to Cursor model names - you choose the base model name and set the thinking level with `shift-tab`, as you normally would. # How it works Each Pi turn spawns a Cursor Agent CLI subprocess. The first turn sends the full Pi transcript; later turns resume the saved Cursor chat session and send only the newest user prompt: ```bash agent --print --yolo --output-format stream-json --model --trust --workspace "" # later turns agent --print --yolo --output-format stream-json --model --resume --trust --workspace "" ``` Cursor session IDs are saved and associated with Pi session IDs. Thanks to that, even if you resume the Pi session, the provider does not need to send the full conversation to Cursor CLI again, but can resume the Cursor session too. The CLI's NDJSON stdout is read line by line, and its output is mapped, when possible, to Pi native rendering elements (for example, thinking blocks). Cursor CLI supports images when you provide a file path to them, so images pasted into Pi are supported out of the box, as Pi provides the path to a temporary file when you paste an image. In non-interactive mode, if the prompt contains a blob of image data, that data is saved as a temporary file, and the path to that file is then passed to the Cursor CLI. # Installing and enabling MCP tools in Cursor Agent for Pi To use Pi-related MCP tools (e.g. `pi-auto`) when the Cursor Agent runs on behalf of Pi, connect the MCP server, enable it for the agent, and allow its tools in the CLI config. ### 1. Connect MCP server to agent Add the server to `~/.cursor/mcp.json`. Example for `pi-auto`: ```json { "mcpServers": { "pi-auto": { "command": "pi-auto-mcp", "lifecycle": "keep-alive", "directTools": true } } } ``` ### 2. Enable the MCP server List MCP servers; new ones need approval: ```bash agent mcp list # Example output: pi-auto: not loaded (needs approval) ``` Enable and approve the server: ```bash agent mcp enable pi-auto # Example output: ✓ Enabled and approved MCP server: pi-auto ``` Verify that the tools are available: ```bash agent mcp list-tools pi-auto ``` Example output: ``` Tools for pi-auto (8): - pi_get_priority () - pi_get_provider (scope, projectPath) - pi_get_strategy () - pi_get_usage (period) - pi_set_priority (priority) - pi_set_provider (provider, model, scope, projectPath) - pi_set_strategy (strategy) - pi_suggest_provider (period) ``` ### 3. Allow tools from this MCP Ensure `~/.cursor/cli-config.json` allows the MCP tools. For example: ```json "permissions": { "allow": [ "Shell(ls)", "Mcp(pi-auto:*)" ], "deny": [] } ``` `Mcp(pi-auto:*)` lets the agent use any tool from the `pi-auto` server. # Possible improvements ### Delegating tool calls to Pi I do not know if this would actually be an improvement, but it is an idea I saw in the [rchern/pi-claude-cli](https://github.com/rchern/pi-claude-cli/) extension: when the CLI outputs a tool call, you kill the CLI, execute the tool call natively with Pi, and then resume the CLI session while providing the tool call results. This would allow the use of Pi native tools instead of Cursor's built-in ones. I do not know, though, how well that would work in practice. # Limitations ### Context/token counts are estimates Cursor CLI does not report exact token/context usage at the end. The provider estimates usage from the serialized prompt and streamed Cursor output so Pi can still show an approximate context percentage. # Troubleshooting #### I don't see any Cursor models Your Cursor CLI executable (`agent`) was not found. Make sure you have it in `PATH`, or set the `CURSOR_AGENT_PATH` or `AGENT_PATH` environment variable to the path of the `agent` executable. # Alternatives - [pi-frontier/pi-cursor-agent](https://github.com/sudosubin/pi-frontier/tree/main/pi-cursor-agent) - most complete implementation that uses the reverse-engineered API. # License ``` MIT License Copyright (c) 2026 Andrey Copyright (c) 2026 Adrian Kepka Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions: The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software. THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. ```