Cargo CLI — AI

Agent resource management: creating and configuring agents, attaching knowledge for retrieval-augmented generation (RAG), connecting MCP servers, and managing agent memories.

For using agents (sending messages, multi-turn chat, polling), use cargo-orchestration. For uploading knowledge files and building knowledge libraries (the content domain), use cargo-content. This skill covers how that knowledge attaches to an agent. For workspace administration — folders (used to organize agents and files), users, API tokens, roles, and submitting reports when the CLI fails — use cargo-workspace-management.

See references/response-shapes.md for full JSON response structures. See references/troubleshooting.md for common errors and how to fix them. See references/examples/agents.md for agent CRUD and configuration examples. See references/examples/mcp-servers.md for MCP server creation and management examples.

Bootstrap

Already signed in (cargo-ai whoami returns a workspace)? Skip to the next section.

npm install -g @cargo-ai/cli            # no global install? prefix every command with `npx @cargo-ai/cli`
cargo-ai login --email you@company.com  # emailed code, no browser; creates the account on first use
                                        # alternatives: --oauth (browser) · --token <api-token> (CI)
cargo-ai whoami                         # confirm the active workspace before any write

Every command prints JSON to stdout; failures exit non-zero with {"errorMessage": "..."}. Anything that creates a run or a batch is async — pass --wait-until-finished or poll the matching get. When the full skill bundle is installed, ../cargo/references/prerequisites.md adds the CLI version pin, token scopes, and the admin-only surface.

Discover resources first

cargo-ai ai agent list                     # all agents (uuid, name, description)
cargo-ai ai template list                  # all AI agent templates (slug, name)
cargo-ai ai mcp-server list                # all MCP servers (uuid, name)
cargo-ai ai memory list --scope agent --agent-uuid <uuid>  # agent memories
# Knowledge files & libraries live in the content domain — see cargo-content:
#   cargo-ai content file list   /   cargo-ai content library list

Retrieve in the UI: agents live at app.getcargo.io/workspaces/<WORKSPACE_UUID>/agents/<AGENT_UUID>. Get <WORKSPACE_UUID> from cargo-ai whoami under workspace.uuid.

Quick reference

cargo-ai ai agent list
cargo-ai ai agent get <agent-uuid>
cargo-ai ai agent create --name <name> --icon-color blue --icon-face 🤖
cargo-ai ai agent update --uuid <agent-uuid> --name <name>
cargo-ai ai agent remove <agent-uuid>
cargo-ai ai release list --agent-uuid <uuid>
cargo-ai ai release get <release-uuid>
cargo-ai ai release get-draft --agent-uuid <uuid>
cargo-ai ai release update-draft --agent-uuid <uuid> --language-model-slug gpt-4o
cargo-ai ai release deploy-draft --agent-uuid <uuid>
cargo-ai ai template list                  # full detail; there is no `template get`
cargo-ai ai mcp-server list
cargo-ai ai mcp-server get <mcp-server-uuid>
cargo-ai ai mcp-server create --name "Internal Tools"
cargo-ai ai mcp-server update --uuid <mcp-server-uuid> --name "Updated Name"
cargo-ai ai mcp-server remove <mcp-server-uuid>
cargo-ai ai mcp-client connect --name "My MCP" --url https://mcp.example.com/sse
cargo-ai mcp                               # serve the platform MCP over stdio
cargo-ai mcp --server <mcp-server-uuid>    # serve a curated workspace MCP server instead
cargo-ai ai memory list --scope agent --agent-uuid <uuid>
cargo-ai ai memory update --mem0-id <id> --scope agent --agent-uuid <uuid> --content "Updated memory"
cargo-ai ai memory remove --mem0-id <id> --scope agent --agent-uuid <uuid>

Agents

Agents are AI resources with configured instructions, a language model, actions, and optional resources.

Before creating an agent from scratch, check existing templates — they capture proven patterns for common use cases (lead research, classification, email drafting) and give you a ready-made system prompt, model, and temperature to start from:

cargo-ai ai template list          # browse patterns — full detail, not a summary
# there is no `template get`: `list` already returns systemPrompt, temperature,
# languageModelSlug, actions and resources, so select the one you want
cargo-ai ai template list | jq '.templates[] | select(.slug == "<slug>")' 

# List all agents
cargo-ai ai agent list

# Get a single agent (includes deployed release details)
cargo-ai ai agent get <agent-uuid>

# Create an agent
cargo-ai ai agent create \
  --name "Lead Researcher" \
  --icon-color blue --icon-face 🤖 \
  --description "Researches leads and enriches data"

# Update an agent
cargo-ai ai agent update --uuid <agent-uuid> \
  --name "Senior Lead Researcher" \
  --description "Updated description"

# Move to a folder (find folder UUIDs via cargo-workspace-management)
cargo-ai ai agent update --uuid <agent-uuid> --folder-uuid <folder-uuid>

# Remove an agent
cargo-ai ai agent remove <agent-uuid>

Agent icon: --icon-color must be one of: grey, green, purple, yellow, blue, red. --icon-face is an emoji string.

Folders: Folder creation, listing, and management lives in cargo-workspace-management (cargo-ai workspaceManagement folder list/create/...). Use that skill to discover or create the <folder-uuid> you pass to --folder-uuid here.

Releases

Releases are versioned snapshots of an agent's configuration (system prompt, actions, resources, model, temperature). Agents execute against their deployed release.

# List releases for an agent
cargo-ai ai release list --agent-uuid <uuid>

# Get a specific release
cargo-ai ai release get <release-uuid>

# Get the current draft release (editable)
cargo-ai ai release get-draft --agent-uuid <uuid>

# Update the draft release
cargo-ai ai release update-draft --agent-uuid <uuid> \
  --system-prompt "You are a lead research assistant..." \
  --language-model-slug gpt-4o \
  --temperature 0.3 \
  --max-steps 10

# Deploy the draft release (makes it live)
cargo-ai ai release deploy-draft --agent-uuid <uuid> \
  --integration-slug openAi \
  --language-model-slug gpt-4o \
  --actions '[]' \
  --mcp-clients '[]' \
  --resources '[]' \
  --capabilities '[]' \
  --suggested-actions '[]' \
  --description "Added research actions"

--with-reasoning [true|false] on update-draft / deploy-draft turns reasoning mode on (bare flag = true) for models that support it.

Coding agents (--harness-slug, --repository)

A release with a harness runs as a coding agent: each turn works in a sandbox with a git checkout instead of a plain chat loop. Harness slugs: claudeCode, codex, openCode, deepAgents.

cargo-ai ai release deploy-draft --agent-uuid <uuid> \
  --harness-slug claudeCode --integration-slug anthropic --language-model-slug <model-slug> \
  --repository '{"connectorUuid":"<github-connector-uuid>","repository":"acme/app","defaultBranch":"main","rootDirectory":null}' \
  --actions '[]' --mcp-clients '[]' --resources '[]' --capabilities '[]' --suggested-actions '[]'

• --repository (JSON): connectorUuid (a GitHub connector — it identifies the user, it is not the identity the agent acts as), repository as owner/name, defaultBranch, rootDirectory, env ([{key, type:"public", value}] or {key, type:"secret", secretValue}), and runAsUserUuid — a service user whose temporary CARGO_API_TOKEN the sandbox gets (the help omits it; see ../cargo-workspace-management/SKILL.md → "Service users"; your permissions must cover it or the call fails 403 runAsUserNotFound / runAsUserNotPermitted).
• The agent acts as the Cargo GitHub app, never as the connected user: it clones, pushes, comments and opens PRs as the app's installation. Install the app on the repo first — without it the turn fails. The sandbox's GitHub token is scoped to that one repository — any other repository is refused — lasts 1 hour, and is refreshed each turn; it is redacted from chat output but readable by the agent.
• It works on a branch cargo/<short-name>, never pushes to the default branch, never force-pushes, and opens or updates the PR through its own PR tool, which links the PR to the chat.
• A mention of the app on a PR the agent's chat opened continues that same chat — no trigger needed (skipped if the chat moved to another repository, or another agent opened the PR). Close the chat to stop it. The GitHub trigger adds allPullRequests (answer mentions on any PR in the repo), fixFailedChecks (+ maxCheckFixAttempts 1–10, reset when someone mentions or reviews) and addressReviews, all default false.

Per chat instead of per release. ai chat create --harness-slug … --repository … and ai message create (for the chat it auto-creates) take the same two flags; omitted, the agent's pinned release decides. They are set when the chat is created: passing either with --chat-uuid is a 400 — harnessSlug and repository are not allowed with chatUuid: they are set when the chat is created. Create a new chat to use another harness or repository.

A message sent to a chat that is still answering is queued: 202 with userMessage and no assistant message yet; the reply follows once the turn ahead finishes (chatClosed → 409, chatNotFound → 404). Sending and polling: ../cargo-orchestration/SKILL.md. In a project: defineAgent({ harness, repository }) — ../cargo-project/references/resources.md.

Structured output & heartbeat — not yet exposed as CLI flags

The release API payload (both draft/update and draft/deploy) accepts two fields that release update-draft / release deploy-draft do not surface as flags (verified against the CLI source — there is no --output / --output-schema or --heartbeat):

The generic --options flag does not carry these — the API's options only holds {connectorUuidsByIntegrationSlug, modelUuidsByIntegrationSlug}. Until the flags ship, set these with a direct API call against the same endpoints the CLI uses:

# Structured (JSON Schema) output on the draft release
curl -sS -X PUT "$CARGO_BASE_URL/v1/ai/releases/draft/update" \
  -H "Authorization: Bearer $CARGO_API_TOKEN" -H "Content-Type: application/json" \
  -d '{"agentUuid":"<uuid>","output":{"type":"jsonSchema","jsonSchema":{"type":"object","properties":{"score":{"type":"number"}},"required":["score"]}}}'
# Deploy carries the same fields — POST .../v1/ai/releases/draft/deploy

draft/update is a partial update: fields you omit keep their current draft values. draft/deploy requires actions, mcpClients, resources, capabilities and suggestedActions in the body. File a workspaceManagement report (see ../cargo-workspace-management/SKILL.md) to request first-class --output / --heartbeat flags — this is the documented feedback channel for CLI/UI parity gaps.

Agent configuration workflow:

1. Browse templates for inspiration: cargo-ai ai template list — it returns each template in full (system prompt, model, temperature, actions), so pick the one closest to your use case straight out of that response
2. Create the agent: cargo-ai ai agent create --name "..." --icon-color blue --icon-face 🤖
3. Get the draft release: cargo-ai ai release get-draft --agent-uuid <uuid>
4. Update the draft with configured actions, resources, prompt, model: cargo-ai ai release update-draft --agent-uuid <uuid> ...
5. Deploy: cargo-ai ai release deploy-draft --agent-uuid <uuid> ...

Templates

Templates are pre-built agent configurations that capture proven patterns for common use cases. Always check templates before designing an agent from scratch — they give you a ready-made system prompt, recommended language model, temperature, and tool configuration that you can adopt as-is or adapt.

# List available agent templates — each entry is complete, so this is the only
# call you need. There is no `template get` subcommand.
cargo-ai ai template list

# Inspect one by slug: filter the same response
cargo-ai ai template list | jq '.templates[] | select(.slug == "<slug>")' 

Templates include a system prompt, actions, resources, and recommended model settings. Use them as a starting point and customize via release update-draft. See references/examples/templates.md for the full guide including an end-to-end example of creating an agent from a template.

Model and temperature guidance

The slug must be on the connector's live list: release update-draft and deploy-draft fail with languageModelNotFound otherwise (checked whenever the release has a connector — the draft's default is the OpenAI one, so an Anthropic slug needs --integration-slug anthropic --connector-uuid <uuid> alongside it). List what a connector offers, with each model's credits per 1,000 tokens and isReasoning, for free:

cargo-ai connection connector get-language-models <connector-uuid>

Price spans three orders of magnitude (gpt-6-luna 0.003 / 1k tokens, claude-sonnet-5 0.2, claude-opus-5-5 1.6, claude-fable-5-1 4), so read it before picking a model for a high-volume agent. claude-3-5-sonnet-latest is marked deprecated in Cargo's Anthropic model list — move agents off it.

Low temperature (0.0–0.2) = deterministic, consistent outputs. High temperature (0.7+) = creative, varied outputs. For production workflows processing thousands of records, prefer low temperature.

Knowledge for RAG (files & libraries)

Knowledge that grounds agent responses (retrieval-augmented generation, RAG) comes from the content domain — see cargo-content:

• Files — uploaded binaries (PDFs, CSVs, text).
• Libraries — collections that group files, either native (workspace-managed) or connector-backed (synced from an external source via an unstructured-data extractor).

Files and libraries moved out of ai into the top-level content domain in CLI ≥ 1.0.19 (cargo-ai content file … / cargo-ai content library …). The old ai file … commands are gone. Everything content-related now lives in cargo-content.

Attaching knowledge to an agent

A file or library is inert until attached to an agent via the draft release's resources array and deployed. Upload files / build libraries in cargo-content, then wire them in here with release update-draft --resources … followed by release deploy-draft. See ../cargo-content/references/examples/files.md for the full upload → attach → deploy sequence.

Only indexed files are searchable. A file stored with indexMode: "none" — the default for content file put, workers, the writeFile node and agent writes — can be listed and read but not searched: capability_file_list marks it isSearchable: false, capability_file_search skips it, and capability_file_analyze refuses it. Attach upload-ed files (indexed by default) for RAG, or re-write the file with --index vectorStore.

Letting an agent write files

A file resource with "isReadOnly": false also grants capability_file_write on the folders it lists ({"kind":"folder","folderUuid":"…"} items; listed files stay read-only). Unset means read-only.

cargo-ai ai release update-draft --agent-uuid <agent-uuid> \
  --resources '[{"kind":"file","slug":"outputs","name":"Outputs","description":null,"prompt":"Save each report here.","isReadOnly":false,"items":[{"kind":"folder","folderUuid":"<folder-uuid>"}]}]'

• The agent writes text it produced, ≤ 1,000,000 characters, to a relative path inside one of those folders (subfolders are created) — nowhere else in the workspace. With several writable folders it must name one.
• Every agent write is stored private, indexMode: none and mirrorToContext: false, even when it overwrites a file someone made public. It returns { fileUuid, path, size, url }, where url is a 1-hour signed link anyone holding it can open.
• A write overwrites the newest file of that name in that folder in place. Give the agent a dedicated output folder rather than one holding knowledge it reads.
• In a project: defineAgent({ files: [{ items: [folder], readOnly: false }] }) — see ../cargo-project/references/resources.md.

MCP — two directions, don't mix them up

MCP (Model Context Protocol) runs both ways in Cargo, and the two surfaces are unrelated:

Before building one, check whether the platform MCP already covers it. Cargo now serves a first-party MCP at https://mcp.getcargo.io/mcp — every workspace member, nothing to deploy — with a small fixed toolset for operating the workspace (whoami, get_usage, search_actions, get_action_schema, autocomplete_action, execute_action, execute_action_batch, get_run, get_batch, list_runs, list_models, describe_model, query_models, plus the read-only context tools map_context, search_context, browse_context, read_context). Hosted clients (ChatGPT connectors, Claude.ai, Cursor over HTTP) point at that URL and sign in with OAuth; the consent screen picks the workspace when the user belongs to several. ai mcp-server is for the other job: a curated, named subset — this tool, that agent, this filtered model — for a client that should see exactly that and nothing else.

Publishing a workspace MCP server

cargo-ai ai mcp-server list
cargo-ai ai mcp-server create --name "CRM tools" \
  --actions '[{"slug":"<action-slug>","kind":"tool","toolUuid":"<tool-uuid>","templateSlug":null,"waitUntilFinished":true,"name":null,"description":null,"isBulkAllowed":false,"config":{}}]' \
  --resources '[{"kind":"model","slug":"<slug>","name":"Accounts","description":null,"integrationSlug":"hubspot","modelUuid":null,"filter":null,"selectedColumnSlugs":null,"limit":null,"prompt":null,"isReadOnly":true}]'
cargo-ai ai mcp-server update --uuid <mcp-server-uuid> --name "Updated name"
cargo-ai ai mcp-server remove <mcp-server-uuid>

• Actions take kind: "tool" or kind: "agent" — an agent can be exposed as a callable MCP tool, not just a tool. waitUntilFinished controls whether the call blocks on the run.
• Resources take kind: "model" (a filtered, column-selected view of a model — keep isReadOnly: true unless the client is meant to write) or kind: "file" (workspace files by UUID, see ../cargo-content/SKILL.md).
• Capabilities (--capabilities, CLI ≥ 1.0.86) expose Cargo's own built-in tools on the server, alongside your actions and resources. JSON array of {slug, config}:

  cargo-ai ai mcp-server create --name "Research" \
    --capabilities '[{"slug":"webSearch","config":{}}]'
  

  Copy
  The eleven slugs are sandbox, memory, context, app, document, webSearch, action, model, file, documentationSearch, and suggestedActions — the same set an agent release takes in its own --capabilities, which is why the examples above pass '[]' rather than omitting it. In a project the same field accepts a bare slug (capabilities: ["webSearch"]).
• update replaces --actions / --resources / --capabilities wholesale rather than merging — read the current server with mcp-server get <uuid> and pass the full array back.

Serving it to a coding agent — cargo-ai mcp

Either server reaches any stdio MCP client through the CLI, using the credentials already on the machine. No token is copied into client config.

claude mcp add cargo -- cargo-ai mcp                     # the platform MCP (no setup)
cargo-ai ai mcp-server list                              # find a curated server's UUID
claude mcp add cargo -- cargo-ai mcp --server <uuid>     # that curated server instead
# Cursor, Windsurf, and other stdio clients: same command as the server entry

With no --server, the bridge uses CARGO_MCP_SERVER_UUID when set, otherwise the platform /mcp. This changed: older CLIs resolved "the workspace's only MCP server" and failed with InvalidUsage when the workspace had none or several — a bare cargo-ai mcp now always has something to serve. stdout carries the MCP protocol and all logs go to stderr, so never print anything to stdout around it.

When to reach for this instead of the skills: the skills give an agent the whole CLI; an MCP surface gives it a bounded set with no shell. Use the bridge for in-conversation lookups and one-off actions, and the CLI for batches, workflows, schema changes, and anything with a cost gate. Full routing rule: ../cargo/SKILL.md → "These skills vs Cargo's MCP surfaces".

Consuming an external MCP server

cargo-ai ai mcp-client connect --name "My MCP" --url https://mcp.example.com/sse
cargo-ai ai mcp-client connect --name "My MCP" --url https://mcp.example.com/sse \
  --disabled-tool-slugs "dangerous_tool,other_tool"

--authentication takes {"issuedAt": "...", "accessToken": "..."} or "null". Connected clients are attached to an agent through its release: release update-draft --mcp-clients …, then release deploy-draft.

Memories

Memories are pieces of information an agent stores from conversations for future reference. They are scoped to a user or to a specific agent (--scope user|agent; there is no workspace scope).

# List agent memories
cargo-ai ai memory list --scope agent --agent-uuid <uuid>

# List user-scoped memories
cargo-ai ai memory list --scope user

# Update a memory
cargo-ai ai memory update \
  --mem0-id <id> \
  --scope agent --agent-uuid <uuid> \
  --content "Updated memory content"

# Remove a memory
cargo-ai ai memory remove \
  --mem0-id <id> \
  --scope agent --agent-uuid <uuid>

Chats, documents, votes, and prompt helpers

Sending messages and polling them is cargo-orchestration's job; these are the housekeeping commands around a chat, plus two prompt-authoring helpers.

# Chats — list (filter by agent, release, user, or how the chat started), read, rename, delete
cargo-ai ai chat list --agent-uuid <uuid> --trigger-type schedule --limit 20
#   --trigger-type: chat | connector | draft | web_extension | app | schedule | workflow
cargo-ai ai chat get <chat-uuid>
cargo-ai ai chat update --uuid <chat-uuid> --name "Q4 pipeline review"
cargo-ai ai chat remove <chat-uuid>

# Feedback on one assistant message, and the votes on a chat
cargo-ai ai vote upsert --chat-uuid <chat-uuid> --message-uuid <message-uuid> --is-upvoted false
cargo-ai ai vote list --chat-uuid <chat-uuid>

# Documents — the versioned text / code / image / sheet artifacts an agent writes in a chat
cargo-ai ai document get <document-uuid>                        # latest version
cargo-ai ai document list --uuid <document-uuid> --limit 10     # its versions
cargo-ai ai document create --uuid <document-uuid> --kind text --title "Brief" --content "…"   # new version
cargo-ai ai document reset --uuid <document-uuid> --created-at <version-createdAt>             # drop every later version

# Suggested follow-ups for a message (cache with --cache-key to avoid regenerating)
cargo-ai ai suggested-action generate --message-uuid <message-uuid> --agent-uuid <agent-uuid>

# Prompt helpers
cargo-ai ai prompt optimize --prompt "find companies in SF and enrich them"
cargo-ai ai prompt evaluate --prompt "<system prompt with {{…}} expressions>" --actions '[...]' --resources '[...]'

• prompt optimize rewrites rough instructions into a structured system prompt (role, behavior, avoidances) with a fixed small model at temperature 0. Treat the result as a draft to edit, not a prompt to ship.
• prompt evaluate does not score anything (its help said it did before CLI 1.0.128). It renders the prompt: template expressions in it are resolved against the --actions, --resources and --capabilities you pass, and the response is the prompt text the agent would actually receive. Use it to check that a system prompt's {{…}} references resolve before deploying a release.
• document reset is destructive — every version created after the one you name is deleted. list --uuid first to pick the createdAt.
• A downvote is feedback the workspace keeps; it does not change the agent. Fixing the answer is a release change (prompt, model, resources).

Help

Every command supports --help:

cargo-ai ai agent create --help
cargo-ai ai release update-draft --help
cargo-ai ai mcp-server create --help
cargo-ai ai memory list --help