dsh-cursor

Cursor models inside DeepSeek Harness.

English | 简体中文

dsh-cursor registers a DeepSeek Harness LLM provider route named cursor. The models it offers are the ones your Cursor account can select, and the streams it returns are produced by the Cursor agent SDK. Once installed, Cursor models appear next to every other provider in the Harness model selector.

Unofficial integration. This project is not affiliated with or endorsed by Cursor / Anysphere. Use it only with accounts you are authorized to use, and in line with Cursor's terms of service.

How it works

Cursor does not publish a chat-completions endpoint. It publishes an agent — a local runtime that owns a prompt, a tool surface, and a model selection. The Harness instead asks a provider for one streamed assistant message per request, with the whole conversation passed in.

This plugin bridges those two shapes:

DSH agent loop                 dsh-cursor                        Cursor
────────────────               ──────────                        ──────
llm.stream(request)  ──►  render transcript as one prompt  ──►  Agent.create(tools:['mcp'])
                          Agent.send(prompt)                ──►  model + custom tools
  ◄── text-delta  ◄────     onDelta → DSH chunks               text / thinking deltas
  ◄── tool-call   ◄────     run.cancel() at the tool boundary  tool choice for a custom tool
run the tool natively

• Every request renders the Harness conversation (system prompt, turns, tool calls and their results) into one prompt for a short-lived Cursor agent. A request stays a pure function of the session log, exactly as the Harness expects.
• Cursor never touches your machine. The agent is created with tools: ['mcp'], so Cursor's own file, shell, search, and subagent tools are not registered. The only tools the model can see are the Harness tools for this request, bridged as in-process custom tools.
• When the model asks for one of those tools, the plugin cancels the Cursor run and reports the tool call to the Harness, which executes it with its own permissions, sandbox, and UI. The result comes back on the next request.
• Text and thinking stream token by token through onDelta; token usage is reported on every turn.

Install

From GitHub:

dsh plugin --profile web add github:atticuswx/dsh-cursor

or from a local clone / tarball:

git clone https://github.com/atticuswx/dsh-cursor
dsh plugin --profile web add ./dsh-cursor
# or: dsh plugin --profile web add ./dist/dsh-cursor-0.2.0.tgz

The package declares a DSH bundle patch, so installation mounts the provider automatically. Reload the page (or restart dsh web) and open Settings > Models: a Cursor row appears with its API-key card.

The bundle is not published to npm; install it from GitHub or from a local checkout. npx dsh-cursor login therefore only works from this repo (node bin/cursor-login.mjs login) or after a global git install — pasting a key into the card needs no CLI at all.

API keys

Open Settings > Models > Cursor. The Cursor card manages any number of API keys and chooses which one every request uses:

• Add key — paste a key and give it a label. The first key becomes active automatically.
• The radio button selects the active key. Requests authenticate as that key's account, and the model catalog follows the active key, so switching accounts is one click (reopen the model picker to reload the new account's models).
• Replace key rewrites the stored secret; Delete removes the entry and its stored credential. A label can be edited without touching the secret.
• Key values are sent to the Harness credentials service under generated CURSOR_KEY_* references and are never returned to the page again. The card shows the reference and the last four characters only.
• The key list and the selection live in this plugin's settings, so they survive restarts; the secrets live in $DSH_HOME/.credentials.yaml.

With no managed key the historic chain still applies, in this order:

1. the key saved in the Models page field (credential CURSOR_API_KEY),
2. the CURSOR_API_KEY environment variable,
3. the credential a previous npx dsh-cursor login stored at ~/.cursor/sdk/auth.json.

Create keys at https://cursor.com/dashboard/api. Team admin API keys are not supported by the Cursor SDK; use a user key or a service-account key.

Browser login (single key, no settings changes)

npx dsh-cursor login     # opens Cursor, stores a user API key under ~/.cursor
npx dsh-cursor status    # signed in?
npx dsh-cursor models    # model ids this account can select
npx dsh-cursor logout    # forget the stored login

That stored login is the last fallback; a managed key takes precedence once one exists.

Models

The catalog comes from Cursor.models.list() for the active account — the selected managed key, or the fallback credential when no key is managed — and is cached for ten minutes per credential. Cursor's per-model parameters are mapped onto the Harness reasoning selector when the parameter is thinking-shaped (thinking, reasoning, effort, …).

If discovery cannot run — for example on an air-gapped pilot — list the model ids explicitly in the plugin configuration:

- id: dsh-cursor
  name: dsh-cursor
  config:
    apiKeyEnv: CURSOR_API_KEY
    models:
      - composer-2
      - gpt-5

What you get, and what you give up

Working:

• model selection, streaming text, streamed reasoning, token usage
• the full Harness tool surface (bash, file edits, grep, todos, subagents, MCP tools, …) with Harness permissions, sandboxing, and approval prompts
• session titles, compaction, and any other auxiliary model call

Deliberate limitations:

• The transcript is replayed on every request. Cursor's agent is stateful and the Harness request is not, so each turn is sent as a full conversation snapshot. Prompt caching keeps this cheap in practice, but reported input tokens are higher than a native provider's.
• Tool calls end the Cursor run. The Harness, not Cursor, is the agent that acts; Cursor's planning between tool calls happens one step at a time.
• No image input. Attachments are replaced with a text placeholder, because the bridged prompt path is text-only.
• Auxiliary calls are also agent runs. Session-title generation and compaction cost one Cursor run each.
• Cursor writes its local agent store under ~/.cursor as the SDK normally does; the plugin sets nothing there.

Troubleshooting

Symptom	Fix
No Cursor models in the picker	No key is active. Add one in Settings > Models > Cursor, paste one in that row's key field, or run npx dsh-cursor status / npx dsh-cursor login.
AUTH failures in the session	The active key or fallback credential expired or was revoked. Replace it in Settings > Models > Cursor, or sign in again.
CURSOR_UNAVAILABLE	@cursor/sdk did not load. Reinstall the package; the host must run Node.js 22.13 or newer.
INVALID_REQUEST naming a model	The model id is not offered to this account. Run npx dsh-cursor models.
QUOTA	The Cursor plan has no remaining usage for this model.

Development

npm install
npm test        # node --test, no network and no credentials needed
npm run check   # syntax check every entry point
npm run pack:dist

A fresh clone runs the credential, catalog, transcript, bridge, and settings card tests, and skips the six checks that import the Harness peer packages. To run those too, make the DSH packages resolvable from this checkout; on a machine with a web profile, link the copies the host itself loads so no duplicate module instances appear:

ln -s "$DSH_HOME/profiles/node_modules/@deepseek-ai" node_modules/@deepseek-ai

node scripts/smoke.mjs --stream exercises the adapter against the real SDK (needs a credential).

Because the package is installed into a profile, activating a newly added bundle needs a fresh profile composition: restart dsh web (or use the in-app plugin manager, which applies live).

Layout:

File	Role
lib/index.js	Plugin entry: provider route, config, credentials, model catalog
lib/client.js	Settings > Models > Cursor card: managed keys and the active selection
lib/keys.js	Key registry normalization and the credential-resolution plan
lib/bridge.js	Cursor run → Harness stream chunks, tool-call handoff, cancellation
lib/transcript.js	Deterministic request → prompt rendering
lib/catalog.js	Cursor.models.list() caching, reasoning-parameter mapping
lib/errors.js	SDK failure → Harness failure-code translation
bin/cursor-login.mjs	login / status / models / logout CLI

Compatibility

Targets DeepSeek Harness 0.2.1-rc.2 (engines.dsh), Node.js 22.13+ (required by @cursor/sdk), and @cursor/sdk 1.2.x.

License

MIT