Sling Assist

Get AI help locally to build, run and debug replications, pipelines, API Specs, etc.

sling assist installs Sling skills and MCP into your local AI agent. After setup, it launches your preffered local AI to help you build, run and debug replications, pipelines, API Specs, etc, by automatically providing useful context.

Quick start

sling assist                     # first run: setup; after: greet with choices
sling assist "backfill orders"   # positional ask (skip the card, launch)
sling assist --id <id>           # investigate a failed run
sling assist --resume            # pick a past session
sling assist --resume <id>       # resume that session
sling assist setup               # re-run setup / change agent
sling assist setup --doctor      # health report
sling assist error <sig>         # look up an error signature
sling assist report --id <id>    # review a redacted failure report
sling assist --out -             # print the prompt, do not launch
sling assist --agent claude      # one-run agent override
sling assist --model sonnet      # pass --model to the agent

First run

If no profile exists, sling assist runs setup.

1. Detect CLI agents on $PATH (claude, codex, gemini, cursor, opencode, pi, grok). VS Code is an install target, not a launchable agent.

2. Rank the detected agents. Sling prefers the one that is signed in.

3. Write the profile to ~/.sling/env.yaml under env.SLING_ASSIST.

4. Write skills to ~/.agents/skills/ and wire each detected client.

sling assist setup --doctor lists only agents on $PATH. A missing binary is not a failure.

You need at least one CLI agent on $PATH, or you can let Sling install OpenCode.

After setup, verify:

sling assist setup --doctor

To change the agent or scope later, run sling assist setup.

Each sling assist call refreshes the skill bundle for wired clients. The refresh also removes retired skills. If no skills are installed, the refresh does not install them. Run sling assist setup first.

Daily use

After setup, sling assist with no ask prints a card and waits. It does not launch an agent until you pick an option or type an ask.

Ask

Pass an ask to skip the card and launch:

sling assist "create a replication from postgres to snowflake"
sling assist "add an incremental stream for orders"

The ask is a trailing argument. --name sets the history slug. Default is a slug from the ask, or assist.

Investigate a failed run

sling assist --id 3f9c2a10

After a failure, Sling prints a hint line with the short exec id:

  investigate with AI -> sling assist --id 3f9c2a10

--id accepts any unique id prefix. It loads that failure snapshot, adds the error signature, the sanitized error excerpt, and the run log tail to the prompt, then launches the agent.

Set SLING_ASSIST_HINT=false to suppress the hint line, or set hint_in_errors: false in the profile.

Failure snapshots live under ~/.sling/assist/errors/executions/. Sling keeps at most 100.

Agent and model

sling assist --agent grok --model grok-4
sling assist --resume --model sonnet

--agent and --model apply to this run only. An empty --model uses the agent default. Sling does not rewrite aliases. The agent validates the name.

Cursor uses the agent binary, not cursor.

Print the prompt

sling assist --out -                          # stdout
sling assist --out /tmp/prompt.md "backfill orders"

--out - prints to stdout. Any other value is a file path.

If you already run inside Claude Code, Cursor, or OpenCode, Sling prints the prompt. It does not launch a nested agent. The same applies off a TTY. Those paths do not show the open screen.

--non-interactive skips interactive forms and launches in headless mode. You must pass an ask or --id. With neither, Sling exits with an error.

Flags on sling assist

Flag	Description
--id	Investigate a failure by its id. An id prefix works.
--resume [id]	Resume a past session. Omit the id to pick.
--agent	Override the profile agent for this run only.
--model	Pass --model to the agent for this run only.
--name	History slug override.
--out, -o	Emit the prompt instead of launching: a file path, or - for stdout.
--non-interactive	Skip forms. Launch headless. Requires an ask or --id.

Pass the ask as a trailing argument.

Subcommands

Command	Description
sling assist setup	Install skills and MCP, or reconfigure / uninstall
sling assist error <sig>	Look up guidance for an error signature
sling assist report --id <id>	Compose a redacted issue report from a failed run

Supported agents

Agent	Launch	Notes
claude	claude	Claude Code CLI
codex	codex	OpenAI Codex CLI
gemini	gemini	Gemini CLI
cursor	agent	Cursor CLI (agent, not cursor)
opencode	opencode	OpenCode. Sling can install this as a fallback
pi	pi	Pi agent
grok	grok	Grok CLI
vscode	—	Skills + MCP only. Not a launchable CLI agent

Skills

Setup writes these skills to ~/.agents/skills/:

Skill	Covers
sling	Sling CLI overview and routing
sling-replications	Replication configs
sling-pipelines	Pipeline configs and hooks
sling-build	SQL models and sling build
sling-api-specs	API spec authoring
sling-platform	Platform commands and the platform MCP
sling-python	The Python wrapper
agent-browser	Browser control for API research

Each client gets a redirect file that points into ~/.agents/skills/.