oc CLI

oc is the first-party command-line client for Open Connector. It uses the native /api/v1 API through @open-connector/sdk, prints readable terminal output by default, and offers explicit, stable JSON through --json for agents and scripts.

Install

curl -fsSL https://github.com/okxiaoliang4/open-connector/releases/latest/download/install.sh | sh

The installer downloads a runtime-free binary for macOS or Linux to ~/.local/bin. Windows users can download oc-windows-x64.exe from the GitHub Releases page.

Log in to an instance

Create a project-scoped API key in the SaaS console, then store the API base URL, key, and optional default entity ID:

printf '%s' "$OPEN_CONNECTOR_API_KEY" | oc login \
  --url https://api.openconnector.dev \
  --key - \
  --entity user_123

This writes ~/.config/oc/config.json. login stores an existing project key; it is not an OAuth login flow. For self-hosting, pass your own public server origin.

Discover, connect, and execute

# 1. Verify setup, then find or inspect an exact tool slug
oc doctor
oc search "create a github issue" --toolkit github
oc tools info GITHUB_ISSUES_CREATE SLACK_CHAT_POST_MESSAGE
oc execute GITHUB_ISSUES_CREATE --dry-run --data @issue.json

# 2. Search for a Toolkit and choose an available connection policy
oc link
# A name such as X returns a bounded, relevance-ranked server result
oc link X

# 3. Execute the reviewed input
oc execute GITHUB_ISSUES_CREATE \
  --connected-account conn_... \
  --data @issue.json

For API-key or Basic providers, supply the credential object instead of completing an OAuth redirect:

oc connected-accounts link telegram \
  --auth-method telegram.api_key \
  --credentials @credentials.json

Commands

Use oc <command> --help for every option. Global flags are --json and --entity <id>.

JSON input and output

--data and --credentials accept inline JSON, @path/to/file.json, or - for stdin:

oc tools execute TOOL_SLUG --data @input.json
cat input.json | oc tools execute TOOL_SLUG --data -
oc --json toolkits list | jq '.items[].slug'

Human-readable output is the default. Toolkit and Tool lists show labeled tables, page context, --cursor continuation guidance, and a copyable batch-capable info command. Agents and shell scripts should pass --json explicitly whenever they parse, store, or pipe output. The primary result format does not change based on terminal detection; only transient fetch status on stderr is limited to an interactive TTY.

Primary results go to stdout; authorization URLs, warnings, and errors go to stderr. tools info accepts 1–20 slugs and always returns { tools, errors, summary }, with the complete detail at each tools[slug]. Partial failures return non-zero unless --allow-missing is explicit. An unhealthy doctor and failed commands return non-zero, with structured diagnostics available through --json. Read secrets with - from stdin or @file where supported. Preview tool writes with --dry-run, and preview connection deletion before confirming it with --force.

In a Human TTY, oc link searches Toolkits remotely with bounded results, reuses a single enabled Auth Config automatically, and prompts only when multiple connection policies or Auth Methods are available. It opens the hosted connection page unless --no-browser is passed. --json, --no-input, and non-TTY calls never prompt or open a browser. Use --timeout 5m to change the wait or --no-wait to return the pending connection immediately; after a timeout, resume with oc connected-accounts wait <id> instead of creating another link.

Agent skills

oc skills install codex
oc skills install claude

The command installs the checked-in CLI instructions into the selected agent's user-level skill directory, honors CODEX_HOME, and refuses to overwrite an existing Skill unless passed --force.

Successful Skill installs are recorded with their exact paths in ~/.config/oc/installations.json. oc skills refresh updates every recorded copy from the current binary, and oc upgrade automatically repeats that refresh after replacing the binary.

Agent Plugins

The separate Open Connector Agent Plugin is a portable Agent Plugins v1 package with a Claude marketplace wrapper.

oc setup
oc setup codex claude --dry-run
oc setup codex claude hermes openclaw
oc setup --refresh

With no arguments, oc setup detects clients and opens a multi-select containing only hosts that expose automatic installation. Manual-only detected hosts are reported separately. Scripts and --json callers must pass explicit agent names; the CLI never prompts without a TTY.

The supported matrix is Claude, Codex, Hermes, OpenClaw, VS Code, Cursor, Grok Bot, Kiro, GitHub Copilot, and NanoClaw. The Agent Plugins specification defines the package and discovery contract, but deliberately leaves installation to each client. oc setup automates verified host CLIs and returns an official setup link for clients without a local installation API; those manual cases return a successful manual result with host-specific steps, but are not recorded as completed installs.

Current boundary

The CLI covers toolkit, tool, and realtime Trigger discovery; auth configs; connected accounts; Trigger Instance control; execution; credential proxying; and a bundled agent Skill. oc search performs deterministic native catalog matching, not semantic search. A realtime Trigger listener/forwarder, outbound webhook-subscription management, and an inline JavaScript sandbox are not part of the current CLI surface.