@rebyteai/cli

Manage saved Agents through the official OpenAI client configured for Rebyte. Install the release in your project:

pnpm add -D @rebyteai/cli

No repository clone is required. The historical ReByteAI/rebyte-cli repository is archived; this package is the current CLI.

export REBYTE_API_KEY='rbk_...'
pnpm exec rebyte agent validate -f /path/to/agent.toml
pnpm exec rebyte agent create -f /path/to/agent.toml
pnpm exec rebyte agent apply agent_... -f /path/to/agent.toml
pnpm exec rebyte agent export agent_... -o /path/to/export.toml

--env dev uses http://127.0.0.1:34567; production defaults to https://api.rebyte.ai. REBYTE_BASE_URL or --base-url accepts the origin or its /v1 URL. --env test requires REBYTE_TEST_BASE_URL. Export refuses to replace an existing file unless --force is present.

Manifest

model = "gpt-5.6-luna"
name = "Company assistant"
instructions_file = "prompt.md"

[[tools]]
type = "mcp"
server_label = "company"
connection_origin = "service"
transport = { type = "http", server_url = "${COMPANY_MCP_URL}" }

[[tools]]
type = "function"
name = "lookup_order"
description = "Look up an order authorized for the current shopper."
parameters = { type = "object", properties = { order_id = { type = "string" } }, required = ["order_id"], additionalProperties = false }

model is required. Use either inline instructions or instructions_file (relative to the manifest). tools defaults to an empty array. Optional settings are reasoning, text, service_tier and metadata, using the Agents configuration fields. Unknown fields and invalid schemas fail before a request is sent. Model availability and provider-specific settings are validated by the server.

Complete ${VARIABLE} tool values are expanded from the environment. Instructions retain literal examples. Do not place HTTP authorization headers in saved Agents; use Session overrides or Vaults. For stdio MCP, omit connection_origin and supply an absolute cwd. Skills, packages, files and network policy belong to the Session.

apply describes the entire saved configuration: omitted optional values clear previous settings. It does not modify existing Sessions. Export preserves native MCP definitions, omits redacted/null optional fields and removes stdio's derived connection origin. Literal JSON null in schemas or request metadata cannot be represented in TOML; export fails without writing. Use the official SDK for JSON.

See migration for retired manifest fields. The Commerce repository has its own local business manifest and Python converter; that file is not an input for this CLI.

Load client functions on demand

Tool Search requires CLI 0.2.1 or later. Include a tool_search entry, then set defer_loading = true on selected function entries:

model = "gpt-5.6-luna"

[[tools]]
type = "tool_search"

[[tools]]
type = "function"
name = "lookup_order"
description = "Look up an order by its ID."
defer_loading = true
parameters = { type = "object", properties = { order_id = { type = "string" } }, required = ["order_id"], additionalProperties = false }

Omitted/false stays eager. Deferred functions without tool_search fail validation. The CLI only manages definitions; your application handles required_actions and returns agent.session.input.tool_result. MCP tools use automatic search independently and need no defer_loading flag.

See the function guide, MCP guide and runnable deferred-functions recipe.

Dynamic Workflow

CLI 0.2.3 and later accepts and exports this singleton tool declaration:

[[tools]]
type = "dynamic_workflow"

Include the MCP or Web Search tools it should orchestrate in the same manifest. Environment tools become available when the application creates a Session with an environment. run_code is reserved and cannot be used as a client function name or MCP server label. See the Dynamic Workflow guide.