openclaw-cursor-cli

An OpenClaw plugin that adds a cursor-cli CLI backend and provider, so OpenClaw can route model calls through the local cursor-agent binary (using your Cursor subscription).

Pattern: structurally identical to OpenClaw's built-in claude-cli backend — a thin wrapper that spawns the Cursor IDE's official CLI in headless -p --output-format stream-json mode and pipes the result back into OpenClaw.

Highlights

• All Cursor models (auto, composer-2, GPT-5.x, Claude 4.5/4.6/4.7, Gemini 3.x, Grok, Kimi, …) usable via cursor-cli/<base>.
• OpenClaw's unified --thinking off|low|medium|high|xhigh|max flag is translated at runtime to the corresponding Cursor model id (-thinking-high, -extra-high, etc.). Catalog stays small (~13 user-facing base ids); per-effort variants are derived.
• Model catalog and thinking-level rules are fetched live from cursor-agent models and cached — no hardcoded lists to drift when Cursor adds models.
• Single /cursor-models refresh slash command (in supported channels) or bash scripts/refresh-models.sh (shell) to re-sync after Cursor updates.

Security & permissions — read first

This plugin acts as a thin CLI bridge to cursor-agent and inherits all of cursor-agent's capabilities. Choose a permission profile that matches your trust level before enabling it for real workflows:

• Using this purely as a model provider? Set mode:"ask" + allowTools:false so cursor-agent runs in read-only Q&A and refuses any tool action. See the Recommended safety profiles table below.
• Working in an untrusted workspace? Add sandbox:"enabled" to confine tool actions to cursor-agent's sandbox, and/or use mode:"plan" to prevent file edits.
• Want the same behavior as OpenClaw's built-in claude-cli backend? Leave config empty — but only do so in workspaces you trust, since cursor-agent can then use its own write/shell tools.
• Model calls run under your local cursor-agent login and consume your Cursor subscription quota. Make sure cursor-agent status shows the Cursor account you actually intend to bill.
• Session resume is enabled so multi-turn conversations work. If you switch between sensitive and non-sensitive tasks in the same channel, start a fresh OpenClaw conversation/session for the sensitive one.

Requirements

• OpenClaw >= 2026.5.7

• cursor-agent installed on PATH and logged in:

  cursor-agent login         # opens browser OAuth, once
  cursor-agent status        # should show ✓ Logged in as ...
  cursor-agent -p "say hi"   # smoke test — must return text
  

Install

Via ClawHub (recommended)

openclaw plugins install clawhub:@jeehou/openclaw-cursor-cli

From a local checkout (development)

git clone https://github.com/jeehou/openclaw-cursor-cli.git
openclaw plugins install ./openclaw-cursor-cli

Post-install (both routes)

# 1. Verify it loaded cleanly. Your existing default model is untouched.
openclaw plugins doctor             # "No plugin issues detected."
openclaw plugins inspect cursor-cli # cli-backend + text-inference both registered
openclaw models                     # Default unchanged

# 2. Fetch the live Cursor catalog, family rules, and auto-allowlist all
#    cursor-cli/<family> entries in agents.defaults.models. REQUIRED — without
#    this step the `/model` picker and `openclaw models` will not show any
#    cursor-cli model.
bash ~/.openclaw/extensions/cursor-cli/scripts/refresh-models.sh

After step 2:

• openclaw models list --provider cursor-cli shows every family with the correct context window (1024k / 256k / 200k).
• /model (slash command in chat channels) lists every cursor-cli/<family> as a selectable option.
• Re-running step 2 is safe: it only adds missing entries to agents.defaults.models and never removes or overwrites entries you (or other plugins) already put there.

Usage

Per-call (recommended — zero default change)

openclaw agent --local --agent main \
  --model cursor-cli/claude-opus-4-7 \
  --thinking high \
  --message "design a worker pool" \
  --json

The --thinking flag is automatically translated:

(Mapping for every family is in ~/.openclaw/extensions/cursor-cli/model-rules.json after refresh. View with bash scripts/refresh-models.sh --dry-run if curious.)

Make Cursor your default

openclaw models set cursor-cli/claude-opus-4-7
# revert any time:
openclaw models set deepseek/ali-deepseek-v4-pro

Refreshing when Cursor adds new models

Whenever Cursor ships new model ids, run one of:

Shell:

bash ~/.openclaw/extensions/cursor-cli/scripts/refresh-models.sh

The script auto-locates your openclaw npm package so the refresh helper can reuse OpenClaw's plugin-sdk subprocess helper (no child_process import on our side). If auto-detection fails — e.g. unusual nvm / yarn layout — override with:

OPENCLAW_PACKAGE_ROOT=/path/to/lib/node_modules/openclaw \
  bash ~/.openclaw/extensions/cursor-cli/scripts/refresh-models.sh

Chat slash command (in channels that support plugin slash commands — telegram / discord / feishu / mattermost):

/cursor-models refresh

This runs the same logic in-process inside the plugin runtime — no shell wrapper, no env-var tricks.

Both routes:

1. Run cursor-agent models and parse all 100+ Cursor model ids.
2. Group ids into families (auto, gpt-5.5, claude-opus-4-7, …) and infer per-family suffix patterns automatically.
3. Build a thinking-level → real cursor id mapping for each family.
4. Persist the catalog to models.providers.cursor-cli (your OpenClaw config).
5. Persist the family-rules cache to ~/.openclaw/extensions/cursor-cli/model-rules.json (read by the plugin's resolveExecutionArgs hook at runtime).
6. Merge each cursor-cli/<family> into agents.defaults.models so the /model slash command and OpenClaw agent allowlist see them. Only adds missing entries — existing entries (yours or other plugins') are left untouched.

Idempotent — re-running just updates the same files. OpenClaw backs up openclaw.json to .bak before every config write.

Other /cursor-models sub-commands (slash only)

Plugin config (optional)

All fields are optional. Defaults preserve full agent mode (backward compatible with 0.0.1). In ~/.openclaw/openclaw.json:

{
  plugins: {
    entries: {
      "cursor-cli": {
        enabled: true,
        config: {
          // Path or basename of the cursor-agent binary (default: cursor-agent on PATH).
          // Only an absolute path or a simple basename like "cursor-agent" is accepted;
          // shell metacharacters (`;`, `|`, `&`, `` ` ``, `$`, `<`, `>`, …) are rejected.
          command: "~/.local/bin/cursor-agent",

          // Execution mode. Pick the lowest privilege your workflow needs.
          //   "agent" (default) — full agent, can plan + edit + run shell tools
          //   "plan"            — read-only/planning, proposes diffs but never edits
          //   "ask"             — pure Q&A, no tool use at all
          mode: "agent",

          // Controls whether cursor-agent may run its write/shell tools without
          // interactive approval. Leave it at the default to keep the standard
          // OpenClaw cli-backend headless behavior (matches `claude-cli`). Set
          // to false in untrusted workspaces — every tool action will then
          // require explicit confirmation. Pairs naturally with mode:"plan"/"ask".
          allowTools: true,

          // Optional override for cursor-agent's sandbox setting.
          //   "enabled"  — confine tool actions to a sandbox.
          //   "disabled" — opt out (cursor-agent's own default behavior).
          //   omitted    — leave it to cursor-agent's config.
          // sandbox: "enabled",

          // Advanced: extra args appended verbatim. Most users do not need this.
          extraArgs: []
        }
      }
    }
  }
}

Recommended safety profiles

Pick the lowest privilege your workflow actually needs.

Upgrading from earlier versions

If you previously installed 0.0.1 – 0.0.4, the upgrade is safe and mostly hands-off. Default plugin behavior is unchanged (omitting mode/allowTools/sandbox keeps the original 0.0.1 semantics) and nothing outside cursor-cli/* in your config is touched.

Run two commands after openclaw plugins install clawhub:@jeehou/openclaw-cursor-cli --force:

openclaw doctor --fix                                                # 1) replay config migrations
bash ~/.openclaw/extensions/cursor-cli/scripts/refresh-models.sh      # 2) re-sync the live catalog

openclaw security audit will also report a cursor-cli.cache.* finding (missing / legacy / stale) when the model-rules cache predates 0.0.5 or is older than 30 days, so you can run the refresh on demand instead of on a schedule.

Uninstall / Rollback

# Light: disable, keep installed
openclaw plugins disable cursor-cli

# Full uninstall
openclaw plugins uninstall cursor-cli

# Also drop the provider config, the rules cache, and the allowlist entries
# the refresh script added:
openclaw config unset 'models.providers.cursor-cli'
rm -f ~/.openclaw/extensions/cursor-cli/model-rules.json
# Strip cursor-cli/* keys from agents.defaults.models (other entries preserved):
PRUNED=$(openclaw config get 'agents.defaults.models' \
  | python3 -c 'import json,sys; d=json.load(sys.stdin); print(json.dumps({k:v for k,v in d.items() if not k.startswith("cursor-cli/")}))')
openclaw config set 'agents.defaults.models' "$PRUNED"

Other providers (deepseek, modelhub, openai-codex, …) are never touched by any of this.

Architecture

openclaw agent --local --model cursor-cli/claude-opus-4-7 --thinking high --message "…"
    ↓
OpenClaw resolves model ref → provider=cursor-cli
    ↓
Routes to CLI backend "cursor-cli" (registered by this plugin)
    ↓
plugin.resolveExecutionArgs reads ~/.openclaw/extensions/cursor-cli/model-rules.json,
   looks up rules.families["claude-opus-4-7"]["high"] → "claude-opus-4-7-thinking-high"
    ↓
Invokes cursor-agent in headless print mode with the stream-json output format
(plus optional flags derived from plugin config: execution mode, sandbox setting,
non-interactive tool approval, session resume) and the resolved Cursor model id.
    ↓
Parses NDJSON stream via OpenClaw's claude-stream-json dialect:
   {type:"system",subtype:"init",session_id}     → session bind
   {type:"thinking",subtype:"delta",text}         → (dropped — Claude dialect doesn't know)
   {type:"assistant",message:{content:[{text}]}}  → assistant text
   {type:"result",subtype:"success",result,...}   → finalize
    ↓
Returns to OpenClaw as a normal assistant turn (session_id persisted for resume)

No HTTP, no OAuth, no protobuf reverse engineering. All auth, rate limiting, and billing remain with the official cursor-agent CLI.

File layout

openclaw-cursor-cli/
├── package.json
├── openclaw.plugin.json          # manifest (id, providers, cliBackends, commandAliases)
├── index.js                      # definePluginEntry: registerCliBackend + registerProvider + registerCommand
├── src/
│   └── refresh-models.mjs        # parse cursor-agent → infer family rules → write user config + cache
├── scripts/
│   └── refresh-models.sh         # shell entry point (used by install + manual refresh)
└── README.md

Limitations

• Streaming: uses --output-format stream-json --stream-partial-output and the claude-stream-json jsonl dialect (Cursor's NDJSON is event-compatible with Claude Code's). Multi-turn session resume via --resume <id> works.
• thinking events from Cursor are silently dropped by the claude-stream-json parser (Claude's dialect doesn't know that event type). Final assistant and result lines are parsed correctly — only intermediate reasoning text is invisible.
• Token usage counters are not surfaced into OpenClaw's agentMeta.usage under stream-json. Switch output: "jsonl" → "json" in index.js if you need accurate accounting.
• By default the plugin runs cursor-agent in OpenClaw's standard headless cli-backend profile (same shape as the built-in claude-cli backend), which lets cursor-agent use its own write/shell tools inside the current workspace. For untrusted workspaces, switch to the safer profiles described under Recommended safety profiles above (e.g. mode:"ask" + allowTools:false, or sandbox:"enabled").
• Per--fast variants (e.g. gpt-5.5-high-fast) are not exposed in the catalog yet. Add a separate --fast flag axis if needed.
• External plugins can't register top-level openclaw <cmd> CLI subcommands in OpenClaw 2026.5.7 — that's why refreshing from the shell goes through scripts/refresh-models.sh rather than e.g. openclaw cursor-models refresh. The /cursor-models slash command is the in-chat equivalent.