@rahularya01/pi-cursor

Use your Cursor subscription's models — Composer, Claude, GPT, Grok — inside the Pi Coding Agent. pi-cursor plugs in a cursor model provider that talks to Cursor's own backend directly (native Connect/protobuf streaming over HTTP/2), so there's no separate API key to buy and no Cursor CLI process running in the background for every chat turn. If you're already logged into Cursor's app or CLI, it just works — no setup beyond installing the package.

Unofficial integration. This project is not affiliated with or endorsed by Cursor / Anysphere. It uses reverse-engineered wire protocol details shared by community clients (see Attributions). Use it only with an account you are authorized to access, and review its source before granting OAuth permissions. Cursor may change wire protocol endpoints or formats at any time.

Contents

• Requirements
• Install
• Quick start
• Authentication
• Commands
• Models & reasoning effort
• Usage dashboard
• Troubleshooting
• Configuration (advanced)
• Architecture (advanced)
• Development

Requirements

Pi Coding Agent / Pi AI	version 0.80.0 or later
Bun	version 1.4.0 or later — the only supported runtime
A Cursor account	with model access — signed in via the Cursor app, Cursor CLI, or browser login below

Install

pi install npm:@rahularya01/pi-cursor

Then restart Pi (or run /reload) so the new provider is picked up.

Install the latest code straight from GitHub instead of npm:

pi install git:github.com/Rahularya01/pi-cursor

To update later:

pi update npm:@rahularya01/pi-cursor

Quick start

1. Sign in. If Cursor's desktop app or CLI (cursor / agent) is already logged in on this machine, pi-cursor detects it automatically — skip to step 2. Otherwise, run:

   /login cursor
   

   This opens a browser tab to sign in with your Cursor account.

2. Pick a model:

   /model cursor/composer-2
   

3. Start chatting. If anything looks off, run /cursor.doctor — it prints which credential source is active, the current endpoint, and the last error, and is the first thing to check before filing an issue.

Authentication and resolution cascade

pi-cursor automatically resolves credentials using a 4-tier cascade:

1. CURSOR_ACCESS_TOKEN environment variable
2. Pi OAuth credentials store (~/.pi/agent/auth.json via /login cursor)
3. Cursor CLI credentials in macOS Keychain (cursor-access-token / cursor-refresh-token)
4. Cursor IDE local state DB (globalStorage/state.vscdb on macOS, Windows, Linux, or WSL)

/login cursor is preferred over Keychain/IDE harvest so an explicit Pi login is not silently overridden by another Cursor app account on the machine.

Automatic CLI & IDE login detection

If you are logged into the Cursor desktop app or Cursor CLI (cursor / agent), pi-cursor automatically extracts your session credentials so you can start chatting immediately without manual browser login.

On WSL (Windows Subsystem for Linux), pi-cursor reuses the current Windows user's Cursor IDE login (USERPROFILE / USERNAME → /mnt/c/Users/<you>/AppData/...). It does not scan other Windows profiles.

To opt out of Keychain / IDE / WSL credential reuse (OAuth or CURSOR_ACCESS_TOKEN only):

export PI_CURSOR_SYSTEM_CREDENTIALS=0

Deep-link PKCE browser login

When no local credentials exist, running /login cursor initiates browser sign-in:

1. /login cursor opens https://cursor.com/loginDeepControl?...
2. Pi polls https://api2.cursor.sh/auth/poll until authentication completes.
3. Access and refresh tokens are stored in Pi's auth store (~/.pi/agent/auth.json).
4. Tokens are automatically refreshed via https://api2.cursor.sh/auth/exchange_user_api_key.

Use /cursor.doctor to inspect which source is active (tokenSource=cli_keychain, tokenSource=ide_vscdb, tokenSource=pi_oauth, tokenSource=env).

Commands

Command	Description
/login cursor	Sign in to Cursor via browser PKCE and refresh the live model catalog.
/model cursor/<id>	Choose a registered Cursor model.
/cursor.models	List active runtime models, context windows, and effort capabilities.
/cursor.models all	Include tab/chat internal model variants normally hidden from the picker.
/cursor.usage	Display visual TUI usage dashboard (included/auto/API quota bars, reset dates, on-demand spend).
/cursor.doctor	Show sanitized provider diagnostics, active token source, endpoint, and last error.

Models and reasoning effort routing

pi-cursor discovers live account models via GetUsableModels and parameterized metadata. Reasoning effort levels (off, minimal, low, medium, high, xhigh, max) are mapped directly to Cursor's runtime model variants or reasoning parameters.

Public model ID	Context	Thinking	Description / Routing
cursor/composer-2	200,000	Yes	Cursor's agentic model with fast reasoning effort options.
cursor/composer-1.5	200,000	Yes	Fast agent model optimized for code edit turns.
cursor/claude-sonnet-5	200,000	Yes	Anthropic Claude Sonnet via Cursor infrastructure.
cursor/gpt-5.5	200,000	Yes	OpenAI flagship model with parameterized reasoning levels.
cursor/grok-4.5	200,000	Yes	xAI Grok model via Cursor infrastructure.

To restrict which models Pi displays, configure ~/.pi/agent/settings.json:

{
  "enabledModels": ["cursor/composer-2", "cursor/claude-sonnet-5", "cursor/gpt-5.5"]
}

Usage quota and visual TUI dashboard

Running /cursor.usage displays a formatted terminal interface showing your current billing cycle, progress bars for included plan quota, auto/API usage, reset dates, and on-demand spend:

Usage • Pro                                           Resets 5 Aug
Monthly plan and on-demand usage

Category        Current          Usage
Included        13% used         ███░░░░░░░░░░░░░░░░░
  Auto          12% used         ███░░░░░░░░░░░░░░░░░
  API           14% used         ███░░░░░░░░░░░░░░░░░
On-Demand       Disabled
------------------------------------------------------------
On-demand usage is off

View in dashboard: cursor.com/dashboard?tab=usage

Usage statistics are fetched directly from Cursor's native Connect period usage endpoint (POST https://api2.cursor.sh/aiserver.v1.DashboardService/GetCurrentPeriodUsage) using your active access token, with optional fallback to CURSOR_USAGE_SESSION_TOKEN.

Architecture & Wire Protocol

The rest of this README is reference material for troubleshooting, tuning, and contributing — nothing here is required for day-to-day use.

Pi Coding Agent  →  streamSimple (cursor-native)
                      → h2-session.ts (in-process HTTP/2 client)
                      → agent.v1.AgentService/Run (Connect + Protobuf over HTTP/2)

• Transport: Native Connect/protobuf streaming over HTTP/2, in-process via h2-session.ts — no subprocess.
• Infrastructure Context Normalization: Side-channel user messages (context-mode routing, post-compaction <session_state>, and explicit [pi-lens automated … not a user request] notices) are safely normalized into the system prompt so Cursor models stay focused on your primary task.
• Context-Efficient Tools: MCP schemas are compacted without changing callable constraints, and exact conversational-only turns (hi, thanks, etc.) omit tools entirely. Actionable prompts always retain tools.
• Cross-Platform: Tested and fully compatible with macOS, Linux, Windows, and WSL.

Configuration

Everything below is optional — pi-cursor works out of the box. These environment variables exist for tuning timeouts, debugging, and edge-case overrides.

Variable	Purpose
PI_CURSOR_AGENT_URL / CURSOR_AGENT_URL	Override agent base URL (default: https://agentn.us.api5.cursor.sh).
CURSOR_ACCESS_TOKEN	Static access token override.
PI_CURSOR_CLIENT_VERSION	Pin x-cursor-client-version header sent by the HTTP/2 bridge.
PI_CURSOR_SYSTEM_CREDENTIALS	0/false to disable Keychain/IDE credential reuse (default: allow).
PI_CURSOR_RAW_MODELS	Disable effort-suffix model collapse.
PI_CURSOR_PROVIDER_DEBUG	Enable verbose JSONL debug logging.
PI_CURSOR_LIFECYCLE_LOG	Always-on compact lifecycle log path (default: $TMPDIR/pi-cursor-lifecycle.jsonl).
CURSOR_USAGE_SESSION_TOKEN	Optional WorkosCursorSessionToken fallback cookie for /cursor.usage.
PI_OFFLINE	Skip live model discovery entirely; always use the bundled fallback catalog.
PI_CURSOR_CACHE_DIR	Where the model catalog and refresh back-off are cached (default: $XDG_CACHE_HOME/pi-cursor or ~/.cache/pi-cursor). Delete it to force a full rediscovery.
PI_CURSOR_UNARY_BRIDGE	1 forces unary RPCs (model discovery) through the general-purpose bridge transport instead of the dedicated one-shot in-process HTTP/2 client. Diagnostic escape hatch.
PI_CURSOR_STREAM_IDLE_TIMEOUT_MS	Silence safety net: ms with no upstream work before recover/retry/error. Default 180000 (3 min); 0 disables (turns run unbounded). Text/thinking/token deltas, tool-call events, and answered execs/queries reset it; heartbeats only prove the socket and do not hide an unanswered exec. It is paused during tool execution. On timeout, recovery continues from checkpoint even after partial output.
PI_CURSOR_RESUME_IDLE_TIMEOUT_MS	Same silence safety net after tool-result resume. Default 180000 (3 min); 0 disables.
PI_CURSOR_STREAM_IDLE_MAX_RETRIES	Auto-recovery attempts after silence/transport loss. Blind restart is skipped once text/thinking streamed unless a checkpoint is available for continuation. Default 5; 0 disables.
PI_CURSOR_ACTIVE_BRIDGE_TTL_MS	How long a mid-tool bridge stays parked waiting for tool results (default: 1 hour).
PI_CURSOR_H2_CONNECT_TIMEOUT_MS	h2-bridge initial connect kill (default: 30000; 0 disables).
PI_CURSOR_H2_IDLE_TIMEOUT_MS	h2-bridge activity idle kill. Default 0 (disabled). Parent heartbeats reset it when enabled.
PI_CURSOR_SLIM_TOOLS	Compact Cursor MCP tool definitions: concise function purpose, no annotation-only parameter prose, full callable schema constraints preserved. Default on; set 0/false for verbatim schemas.
PI_CURSOR_MIDPAUSE_REBUILD_MAX_AGE_MS	Max age of mid-pause metadata used for full-history rebuild (default: 15 min).
PI_CURSOR_PROMPT_HISTORY	Publish the system prompt and completed turns as Cursor prompt messages when a request is built without an upstream checkpoint. Default on; 0/false restores the pre-1.4.24 behavior, where a rebuilt conversation reached the model with no history and no Pi system prompt.

Architecture notes

Stream modules are split under src/stream/:

Module	Responsibility
types.ts	Shared structural types (no runtime code — safe for anyone to import)
config.ts	Agent URL + client version resolution
tuning.ts	Timeouts, retry budgets, and the stream idle watchdog
debug-log.ts	Debug / lifecycle / metric sinks with secret redaction
images.ts	Image decode + Cursor CLI format and size validation
model-routing.ts	Effort suffix / requested model resolution
model-discovery.ts	GetUsableModels unary RPCs + per-token model cache
model-cache.ts	Cross-process catalog cache read synchronously at startup
context-normalize.ts	Context-mode side-channel folding
message-parsing.ts	Pi/OpenAI message list → Cursor turn structures
pi-adapter.ts	Pi context/model types ↔ OpenAI-shaped request, usage accounting
request-build.ts	AgentRunRequest protobuf construction + blob store
bridge-session.ts	Active-bridge registry + h2-bridge lifecycle
session-state.ts	Conversation store, checkpoints, key derivation, session locks
server-messages.ts	Inbound KV / exec / interaction dispatch
thinking-filter.ts	Strips inline <think>-style tags from the text channel
recovery.ts	Tool-continuation recovery planner
protocol.ts	Auth/protocol error enhancement
drift.ts	Wire-drift detection (unknown message cases and protobuf fields)
native-core.ts	Native streamSimple runtime that drives all of the above

Native streamSimple is the only chat path. The OpenAI-compatible local proxy that used to sit alongside it was removed in favour of a single code path.

Startup

Extension activation does no network and no credential lookup. Models are registered synchronously from the persisted catalog (PI_CURSOR_CACHE_DIR), falling back to the catalog bundled in src/models/catalog.json on a first-ever launch. Live discovery runs through pi's refreshModels hook — off the critical path, in the background, and again whenever /model is opened — then persists its result for the next launch.

All Cursor HTTP/2 transport runs in-process via node:http2, which Bun implements natively — no subprocess is spawned. Unary RPCs (both discovery calls) use a dedicated one-shot client (h2-unary.ts); the bidirectional chat stream uses a persistent session (h2-session.ts) that survives across turns. Unary calls fall back to the general-purpose bridge transport if the one-shot client fails.

src/proto/agent_pb.ts is a large generated Connect/protobuf surface used by the wire layer. Never hand-edit it — regenerate with bun run proto:gen (see proto/README.md) when Cursor changes the agent schema.

Troubleshooting

• No API provider registered for api: cursor-native: Update to the latest pi-cursor (pi update npm:@rahularya01/pi-cursor) and restart Pi (or /reload). This means the Agent tried to stream via Pi's global streamSimple dispatcher before the Cursor transport was registered there. Current builds register cursor-native on that registry during extension load.
• Not logged in / 401: Ensure Cursor CLI or app is logged in, or run /login cursor again. Check /cursor.doctor to verify your tokenSource. Tokens from CLI/IDE are re-resolved when near expiry; idle stream retries also force-refresh credentials.
• Empty / hung stream: Cursor may have updated wire headers; verify network connectivity or bump PI_CURSOR_CLIENT_VERSION. /cursor.doctor prints the active clientVersion.
• Wire-protocol drift: Cursor can change agent.v1 at any time. Unrecognized server messages and unknown protobuf fields are no longer skipped silently — they are counted, written to the lifecycle log as wire_drift, appended to the failing turn's error message, and listed by /cursor.doctor under wireDrift. wireDriftStranding=yes means an unanswered message could have parked the turn, which is the difference between "our schema is a bit behind" and "this is why it hung". Run CURSOR_ACCESS_TOKEN=... bun run smoke:wire to check the handshake and schema against the live endpoint without starting a chat turn, then see proto/README.md to resync the schema.
• Stuck / dies after a few minutes of work: Cursor InteractionQuery prompts are answered so the stream does not park. Web/search and unnamed proto fields are rejected (use Pi tools instead). Inspect $TMPDIR/pi-cursor-lifecycle.jsonl for interaction_query / bridge_close events, and /cursor.doctor for lastStreamEvent. Full debug: PI_CURSOR_PROVIDER_DEBUG=1.
• Tool continuation lost: The provider now prefers full-history rebuild when checkpoints are stale/mismatched. If recovery still skips, /cursor.doctor shows lastRecoverySkipReason. Retry the turn or start a new chat.
• WSL credential detection: Set USERPROFILE or USERNAME so the Windows home directory is known, and ensure /mnt/c/Users/<you>/AppData/... is readable. Disable with PI_CURSOR_SYSTEM_CREDENTIALS=0 if undesired.
• Slow startup: Activation should be a few milliseconds. /cursor.doctor reports catalogCache (none(using bundled fallback) means every launch is starting cold — check that catalogCacheDir is writable) and unaryTransport. A stale Cursor CLI keychain entry no longer blocks startup: a refresh token that fails is remembered for 10 minutes so it is not retried on the next launch, and any valid locally stored token is always preferred over a network exchange.
• Model list looks stale: It is the last successfully discovered catalog. Open /model to trigger a background refresh, or delete PI_CURSOR_CACHE_DIR to force full rediscovery.

Runtime

pi-cursor targets Bun only — no Node.js binary is required or spawned at any point. All Cursor HTTP/2 transport (the bidirectional chat stream and the unary discovery RPCs) runs in-process via node:http2, which Bun implements natively.

Earlier versions proxied the chat stream through a short-lived Node subprocess, because Bun's node:http2 client was believed unable to carry a bidirectional Connect stream reliably. That subprocess is gone: oh-my-pi, a Bun-hosted fork of Pi that talks to the same Cursor RPC, demonstrates the same bidirectional pattern working fine in-process under Bun. Its only documented Bun/H2 caveat is ALPN negotiation failing behind an ALPN-stripping TLS-intercepting proxy (e.g. Zscaler) — an environment issue, not a bidirectional-streaming bug — and /cursor.doctor's lastStderr/lifecycle log will name that explicitly if it happens.

/cursor.doctor reports the runtime as runtime=bun <version>.

Development

The toolchain is Bun — package manager, script runner, test runner, and bundler. tsc still does the typechecking, and ESLint and Prettier are unchanged.

bun install
bun run check

bun run check runs TypeScript typechecking, ESLint, Prettier format verification, security checks, the protobuf staleness check, and unit tests.

Script	Purpose
bun run proto:gen	Regenerate src/proto/agent_pb.ts from proto/agent.proto.
bun run proto:sync	Rebuild proto/agent.proto from an updated generated file obtained upstream.
bun run proto:check	Fail if the generated protobuf is stale or hand-edited (part of bun run check).

Attributions

Wire protocol and authentication patterns adapted from MIT community client lineage:

• ephraimduncan/opencode-cursor
• @pi-stef/cursor

Package structure mirrors pi-antigravity.

Support the project

If pi-cursor is useful to you, consider sponsoring the project on GitHub.

License

MIT