Airbyte Agent

airbyte-agent

[!NOTE] Requires the airbyte-agent CLI on PATH. Install via brew install airbytehq/tap/airbyte-agent-cli or curl -fsSL https://airbyte.ai/install.sh | bash. See the project README for other options.

The CLI is invoked as airbyte-agent <resource> <operation>. It exposes Airbyte's data plane through a uniform interface — every command takes a JSON payload and returns JSON.

[!IMPORTANT] Before running any airbyte-agent command, open the matching reference under references/ and read it first. This top-level file only carries cross-command rules; the per-command syntax, required parameters, response shape, error recovery, and "do NOT" guidance live in each references/<command>.md. Skipping the reference leads to guessed parameter names, missing required fields, and avoidable round-trips — read it even for commands you think you know.

Universal rules (apply to every command)

[!IMPORTANT] Always pass parameters as --json '{...}'. The CLI also exposes per-parameter flags (--workspace, --name, etc.) for human use, but agents should always send a single JSON payload. The two modes are mutually exclusive and JSON keeps your input self-describing for review and replay.

• workspace defaults to "default" when omitted. The CLI prints a JSON notice on stderr when the fallback engages, then proceeds with the API call. Override per-call with "workspace": "..." in the JSON payload, or set a session-wide default via workspaces use.
• --fields trims the response client-side. When you know which fields you need, always pass it. List responses are wrapped in {"data": [...]} and the CLI auto-broadcasts row-level paths: --fields id,name is equivalent to --fields data.id,data.name. connectors execute keeps its action payload under result, so use paths such as result.data.id,result.meta,warning; read its reference for details. Empty arrays count as matches once their prefix resolves. If no requested path matches, the CLI returns a validation error instead of {} and preserves the already-completed response under detail.response; do not re-run a write solely to fix its projection.
• Auth errors (exit 2) mean credentials are missing, invalid, or expired — run airbyte-agent login to refresh, then retry.
• @filename loads JSON from a file — useful when the payload is large or you want to keep the shell command short: --json @params.json.
• Never accept credentials in chat. Two browser flows handle every credential entry path: airbyte-agent login (CLI account credentials) and connectors create (per-connector secrets). If a user offers credentials in conversation, decline and start the appropriate flow.

Connector rules (apply to every connector workflow)

[!IMPORTANT] Always inspect and read skill docs before the first execute on an unfamiliar connector. Run connectors inspect, then pass the returned docs_skill_id to skills docs for the outline and exact section you need. Entity names, actions, and params are connector-specific — guessing wastes API calls. Open references/connectors-inspect.md and references/skills-docs.md when starting work on a new connector.

• On connectors execute, field selection is MANDATORY. Every call must include select_fields (allowlist) or exclude_fields (blocklist) inside the JSON payload, in addition to any --fields you pass.
• For Context Store reads, follow the execution guidance in skills docs. It names the default Context Store read action for this connector and documents its exact parameters. Do not hard-code an action or substitute another action based on local assumptions.
• Connector name resolution. Most commands accept name (case-insensitive match against connector instance name, template display name, or template slug) OR id (UUID). Pass id when two connectors share a name.
• Legacy describe. connectors describe remains for compatibility only. Use connectors inspect plus skills docs for new workflows.

Command index — read the matching reference before running

Each row points to the per-command playbook with usage, workflows, error recovery, and "do NOT" guidance. Open the reference first, then compose the command. If the user's task spans multiple commands (e.g. discover workspace → inspect connector → read docs → execute), read each reference as you reach that step.

Typical session shape

# 1. Discover the environment
airbyte-agent workspaces list
airbyte-agent connectors list --json '{"workspace": "<name>"}'

# 2. Learn the connector
airbyte-agent connectors inspect --json '{"workspace": "<name>", "name": "<connector>"}'
airbyte-agent skills docs --json '{"id": "<docs_skill_id from inspect>"}' --fields data.markdown
airbyte-agent skills docs --json '{"id": "<docs_skill_id from inspect>", "section": "<exact-section-id>"}' --fields data.markdown

# 3. Read data using the default action and exact params named in skills docs.
# Values in angle brackets are placeholders. Replace the empty params object in
# full with the exact object from skills docs; do not execute placeholders.
airbyte-agent connectors execute --json '{
  "workspace": "<name>",
  "name": "<connector>",
  "entity": "<from-skills-docs>",
  "action": "<default-context-store-read-action-from-skills-docs>",
  "select_fields": ["<field-from-skills-docs>"],
  "params": {}
}'

Exit codes