cinna-cli 0.2.3

Local development CLI for Cinna Core agents

cinna-cli

Local development CLI for Cinna Core agents.

Work on agent scripts, prompts, and webapps locally with your own editor and AI tools. The CLI keeps your workspace continuously synced with the remote agent environment, streams commands to it, and wires up MCP integration — so the platform is the single source of truth for runtime and credentials.

How It Works

Cinna Core agents run in managed cloud environments. cinna-cli does not run a local Docker container. Instead:

1. Continuous sync — Mutagen keeps ./workspace bidirectionally synced with the remote agent env over a WebSocket tunnel to the platform.
2. Remote exec — cinna exec <cmd> streams your command through the platform to the remote env, with live stdout/stderr and the remote process's exit code.
3. MCP integration — the local MCP proxy gives Claude Code / opencode access to the agent's knowledge base.

Your Editor / Claude Code
        │
        ▼
  workspace/              ← edit locally
        │
  cinna sync (Mutagen) ◄──► Remote Agent Environment   (no local container)
        │
  cinna exec <cmd>       ── streaming output

Prerequisites

• Python 3.10+
• Mutagen (version pinned by the platform — cinna setup checks and prompts to install)

Getting Started

Setup is initiated from the Cinna Core platform UI. Click "Local Development" on your agent's page to get a bootstrap command:

curl -s https://your-platform.com/api/cli-setup/TOKEN | python3 -

This will:

1. Install cinna-cli (via uv, pipx, or pip)
2. Exchange the setup token for CLI credentials
3. Verify / prompt-install the required Mutagen version
4. Clone the workspace (one-shot tarball; Mutagen takes over afterwards)
5. Generate CLAUDE.md, BUILDING_AGENT.md, .mcp.json, opencode.json, .gitignore, mutagen.yml
6. Start the continuous sync session

After setup:

cd hr-manager-agent/
cinna dev                           # start a foreground dev session (live sync + TUI)
claude                              # open Claude Code (MCP tools auto-configured)
cinna sync status                   # see sync state from another terminal
cinna exec python scripts/main.py   # run a command in the remote env
cinna list                          # see every agent registered on this machine

Commands

cinna setup <token_or_url>

Initialize a local workspace. Accepts the setup token, the URL, or the full curl command from the platform UI.

The agent directory name is normalized to lowercase with dashes ("HR Manager Agent" → hr-manager-agent/).

cinna set-token <token_or_url>

Refresh the CLI token on the current workspace without re-cloning. Run this from inside an existing agent directory when the stored token has expired — cinna set-token re-exchanges the setup token via POST /api/cli-setup/{token} and swaps the result into .cinna/config.json and ~/.cinna/agents.json in place. Workspace files, mutagen.yml, and generated context files are left untouched.

Accepts the same input forms as cinna setup (curl command, URL, or bare token). When only a bare token is given, the platform URL is reused from the workspace's existing .cinna/config.json — so you can refresh each agent from inside its own directory even if different agents live on different platforms. The exchanged token must belong to the same agent as the workspace; mismatched agent IDs abort the refresh.

cd hr-manager-agent/
cinna set-token yWo36tbkdAOzrALxOEKq31_OA2iMelEg

cinna login [domain]

Sign in to an account workspace in the browser — no setup token to paste. One command serves two cases:

• Resume — run it from inside an existing account workspace and it refreshes the stored account CLI token in place (reusing the platform URL + machine name; other settings like the active user workspace are preserved).
• Connect new — run it anywhere else and it bootstraps a fresh account workspace. It asks for the platform domain (or pass it as an argument — protocol optional, localhost recognized), then creates the workspace in the current folder if it's empty, or in a subfolder you name if it isn't. So you can mkdir my-cinna && cd my-cinna && cinna login, or just cinna login app.example.com straight into the current directory.

Either way it opens a browser authorization URL (OAuth 2.0 device flow); once you click Authorize (already signed in to the platform) the CLI receives a fresh token and writes .cinna/account.json.

# Resume the account workspace you're in:
cd my-cinna/ && cinna login

# Connect a new account from a fresh folder:
mkdir my-cinna && cd my-cinna && cinna login          # prompts for the domain
cinna login app.example.com                            # or pass it directly
cinna login app.example.com --dir my-cinna             # always into a named subfolder
#   Your verification code: WX7K-9Q2P
#   Open this URL and click Authorize:
#     https://app.example.com/device?code=WX7K-9Q2P
#   ✓ Account workspace ready.

Use it when cinna account status or cinna doctor reports the account token has expired. Because the per-agent tokens minted from an account (cinna agent sync) can only be re-minted while the account token is valid, the flow is: cinna login (refresh the account), then cinna doctor (re-mint the dependent sub-agent tokens). If the platform doesn't expose the device-login endpoints yet, cinna login says so and points you at the cinna account setup paste fallback.

cinna account setup <token_or_url>

Initialize an account workspace — a multi-agent root from which you can discover your agents and attach per-agent workspaces without going back to the UI. Setup is initiated from Settings → Channels → Local Development on the platform, which emits a curl | python3 one-liner (same pattern as the per-agent flow):

curl -sL https://your-platform.com/api/cli-setup/account/TOKEN | python3 -

Accepts the same input forms as cinna setup (full curl command, URL, or bare token — bare tokens need CINNA_PLATFORM_URL). Creates my-cinna/ (override with --dir) containing:

my-cinna/
  .cinna/account.json    # account CLI token + platform/frontend URLs (0600, do not commit)
  CLAUDE.md              # orchestrator guide for AI tools
  context/               # platform docs, API reference, example scripts, worked playbooks
  agents/                # one standard per-agent workspace per `cinna agent sync`

Setup also downloads the context package into context/ — curated platform docs (feature map at context/platform/README.md), a generated per-domain REST API reference (context/api_reference/), sample platform-API scripts (context/examples/), and worked end-to-end playbooks (context/guides/, e.g. build-an-agentic-network.md). The download is best-effort: if it fails, setup still succeeds with a warning and cinna account refresh-context fetches it later.

The account token is only used for the account-level endpoints (listing agents, minting per-agent tokens, the context package). Per-agent work always runs on each child workspace's own token. Revoking the account session in Settings disconnects every agent synced from it.

cinna account agents

List the agents your account can access (run from inside the account workspace). For each agent: display name + ID, building rights (✓ can build, view-only, or foreign install — installed bundles are publisher-managed and can't be synced), whether a remote environment is active, and whether a local workspace already exists under agents/.

cinna account status

One-shot summary of the account workspace: platform/frontend URLs, machine name, synced-agent count, and an account-token probe (valid token / expired token / no connection) — the account-level counterpart of cinna status.

cinna account refresh-context

Re-download the context package and replace the account workspace's context/ tree (run from inside the account workspace). Use it when the platform ships updated docs or API reference. The existing tree is only removed after a successful download — a failed refresh warns and leaves the previous context/ intact.

cinna account user-workspace list | activate <name|id> | clear

Choose the active user workspace for the account session. Workspace-scoped resources you create from the account workspace — new agents (cinna agent create) and the credentials they acquire — land in the active workspace, just like picking a workspace in the web sidebar before creating things. The selection is stored client-side in .cinna/account.json; the platform keeps no active-workspace state.

cinna account user-workspace list                 # show workspaces, marking the active one
cinna account user-workspace activate Sales       # by name or id
cinna account user-workspace activate default     # clear back to the Default (unassigned) workspace

cinna account credentials list | types | create | update | delete | share-with-agent

Draft and wire the credentials your agents need — without ever handling secret values. The account CLI scaffolds a credential as a draft and attaches it to an agent; the user fills the secret in the web UI (the draft shows as "needs setup" until then). This lets a local coding agent set up everything an agent requires and simply tell the user what to fill in — they don't have to think about what to create or share.

The account token can never read or write a credential's secret value — these verbs touch only metadata and structure.

cinna account credentials types                                   # types + the fields the user must fill
cinna account credentials create --name "Stripe Key" --type api_token \
    --agent billing-agent                                         # create a draft and attach it in one step
#   → prints required fields (e.g. api_token) + a link to fill them in
cinna account credentials list                                    # name, type, status (complete / needs setup)
cinna account credentials share-with-agent <cred_id> --agent crm-agent
cinna account credentials update <cred_id> --name "Stripe (live)"
cinna account credentials delete <cred_id> --yes

A new draft lands in the account's active user workspace. Deletes reuse the platform's blast-radius gate (a publisher-provided credential in a published bundle with active installs needs --force). All write verbs require the agent-developer role.

cinna agent create <name> [--description TEXT]

Create a new agent on the platform from the account workspace — no UI interaction. Thin client: only the name (and optional description) is sent; the backend applies all defaults (default AI credentials, env template, environment creation) exactly as creating from the UI does. The agent is created in the account's active user workspace (if one is set). Prints the created agent's ID and web UI link, plus a hint to attach a local workspace with cinna agent sync <name>. Requires the agent-developer role (403 otherwise). Template selection is not supported yet — agents always get the server default.

cinna agent create "CRM Agent" --description "Tracks customer accounts"
cinna agent sync crm-agent

cinna agent sync <agent>

Mint a per-agent CLI token (no UI interaction) and materialize a standard workspace under agents/<slug>/. <agent> is the display name, slug, or agent ID from cinna account agents. The result is identical to what cinna setup produces — own .cinna/config.json, registry entry, generated CLAUDE.md / BUILDING_AGENT.md / MCP configs / mutagen.yml, and the initial workspace clone — so afterwards:

cd agents/hr-manager-agent/
cinna dev

works exactly as for a manually set-up agent. Synced agents also appear in cinna list and in the agent's Integrations-tab session list like any other CLI session. The backend gates minting on building rights: foreign bundle installs and view-only agents are rejected with the server's error message.

cinna agent unsync <agent>

Detach a synced workspace: stop its sync session, revoke the minted CLI token server-side (via the account-scoped revoke endpoint, authenticated with the account token; idempotent), then perform the equivalent of cinna disconnect: remove .cinna/, generated files, and the registry entry. Workspace files under agents/<slug>/workspace/ are preserved. The revoke degrades gracefully — if it fails (no connection, or a workspace synced before token-id tracking), a warning is printed and the local teardown still completes; the token then expires on its own or can be revoked from the agent's Integrations tab.

cinna connect agent-api --producer <agent> --consumer <agent> [--label TEXT] [--read-only]

Wire one agent to another's REST API from the account workspace. Resolves both agents (name, slug, or ID), mints a producer API token, and attaches it to the consumer as a credential — which rides the consumer's normal credential sync into its remote environment, so no key ever touches your machine. Prints the credential ID, token prefix, base URL, and spec URL. --read-only restricts the consumer to read-only access; --label names the credential. Backend errors surface verbatim: 400 if the producer's REST API is disabled, 403/404 for ownership violations.

cinna connect agent-api --producer crm-agent --consumer hr-manager-agent --read-only

cinna connect mcp --producer <agent> --consumer <agent> [--label TEXT] [--conversation-only|--building-only]

Wire one agent to another's agent2agent MCP connector. The consumer is resolved from your agents; the producer is resolved against the platform's discoverable-connectors listing (it must expose an agent2agent MCP connector your account is allowed to consume — the error lists the discoverable options otherwise). By default the connection is enabled in both conversation and building modes; scope it with --conversation-only / --building-only. Prints the credential ID, endpoint, transport, and status — plus an authorize URL to open if the connector requires OAuth.

cinna connect mcp --producer crm-agent --consumer hr-manager-agent --building-only

cinna api <METHOD> <path> [--json TEXT | --data @file.json] [--query k=v ...]

Generic escape hatch into the platform API, authenticated with the account token (run from the account workspace). <path> is relative to the API root — agents, agents/<id> — no /api/v1 prefix. The endpoint catalogue ships in the account workspace under context/api_reference/.

• --json '<obj>' or --data @file.json supply a JSON request body (mutually exclusive); repeatable --query k=v supplies query parameters (repeating a key builds a list).
• The inner response is passed through verbatim: the body prints to stdout (pretty-printed for JSON) and the exit code is 0 for 2xx and 1 for an inner 4xx/5xx — so it composes in shell pipelines.
• When the escape hatch itself refuses the call, the detail prints to stderr and the exit code is 2: policy denials (credentials, user management, admin, CLI, MFA/auth, and streaming routes are excluded — shown as blocked by platform policy: …), rate limiting (429, with the Retry-After delay), and request/response size caps (413/502).

cinna api GET agents
cinna api GET agents --query limit=5
cinna api PATCH agents/3fa85f64-5717-4562-b3fc-2c963f66afa6 --json '{"description": "updated"}'
cinna api POST tasks --data @task.json

cinna chat [--agent <ref>] [--resume <session_id>] [--file PATH ...] [MESSAGE...]

Talk to an agent through a real platform session — the same conversation pipeline production uses (permission checks, agent-env calls, the model/SDK the platform selects), not a local mock. Built for a local coding agent to test the agent it is building: it can prepare a prompt, attach files, and read the reply back as structured data.

Run it from the account workspace (or any synced agent folder under it). The reply is observed by polling the backend rather than reading a live stream, so it is robust to streaming/transport quirks.

• --agent <name|slug|id> picks the agent; omit it inside a synced agent workspace to infer it. --resume <session_id> continues an existing conversation instead of opening a new one (default mode for a new session is conversation; --mode building opens a building session).
• The message is the positional argument; if omitted it is read from stdin, or you are prompted for it interactively in a TTY.
• --file PATH (repeatable) uploads a local file and attaches it to the message.
• Output is NDJSON by default — one JSON event per line (session, upload, message, status, done), trivially parseable by another agent. --pretty switches to a human-readable transcript.
• Each message carries the agent's reasoning/tool trace under events — an ordered list of the thinking blocks, tool calls (with their full tool_input payload) and tool results behind the reply, so you see what the agent did, not just its final content. Pass --no-events to drop the trace and keep only the final text.
• Files the agent attaches to its replies are downloaded under ./cinna-chat-files/<session_id>/ (override with --download-dir, or skip with --no-download to just report the file ids). Downloads are bounded by the api-proxy's 8 MiB response cap.
• --interval / --timeout tune the poll cadence and the maximum wait for a turn. Ctrl-C interrupts the agent's turn and exits.

cinna chat --agent crm-agent "Summarize today's leads"
cinna chat --agent crm-agent --file report.csv "Validate this export"
cinna chat --resume 3fa85f64-5717-4562-b3fc-2c963f66afa6 "Now break it down by region"
echo "ping" | cinna chat --agent crm-agent          # message from stdin
cinna chat --agent crm-agent "hi" | jq -c 'select(.event=="message")'

The session id is printed in the first session event — capture it to drive a multi-turn conversation with --resume.

cinna dev

Start a foreground dev session — creates / resumes the Mutagen sync session for this workspace and attaches the terminal to a two-tab TUI (status + raw Mutagen details). Ctrl-C terminates the session; sync does not outlive the TUI. To observe sync from another terminal without affecting it, use cinna sync status.

cinna redev

Like cinna dev, but conflicts surfaced by the initial reconciliation are resolved automatically in favor of the remote version. Use it to resume work on an agent that was modified from the platform side while your local copy sat idle — same connection, same credentials, no re-setup; the remote data simply wins the startup divergence.

The displaced local versions are backed up under .cinna/sync/redev-backup/<timestamp>/ before being overwritten. Only startup conflicts are auto-resolved — conflicts that arise later in the session are surfaced normally (Conflicts tab / cinna sync conflicts).

cd hr-manager-agent/
cinna redev    # remote wins the initial conflicts, then a normal dev session

cinna sync status | conflicts | push | pull | resolve

Inspect and drive the sync session. status / conflicts are read-only views (safe alongside a live cinna dev); push / pull / resolve are for scripted (headless) builders who aren't running the TUI. All accept --agent <ref> to target a synced child workspace from the account root.

• status — state, pending changes, conflict count. Warns loudly when conflicts mean your edits aren't fully live.
• conflicts — list conflicted paths (sourced from the Mutagen daemon, so it agrees with status; two-way-safe writes no .conflict.* files on disk).
• push [--force] — ensure a session, then flush and block until settled. --force resolves any parked conflicts in favor of local first ("my local is the truth"). The session persists in the daemon so later edits keep syncing.
• pull [--force] — the mirror; --force resolves in favor of remote (e.g. after the backend regenerates managed files).
• resolve --prefer local|remote — clear parked conflicts in one command. local deletes the remote losing copies (your version propagates out); remote backs up your local copies under .cinna/sync/ and takes the container's version. Replaces the manual kill/delete/restart dance.

cinna exec <command…>

Stream a command through the platform to the remote agent environment. Output streams back live; Ctrl+C aborts. Exit code matches the remote process.

The command runs with the workspace root (/app/workspace) as its working directory, so relative paths resolve against the synced workspace — e.g. cinna exec python scripts/main.py runs /app/workspace/scripts/main.py (the same cwd the scheduler uses). No need to prefix paths with /app/workspace/.

Arguments pass through transparently — each token is re-quoted before being sent, so spaces and shell metacharacters inside an argument survive intact. Use ordinary single-level quoting, exactly as for a local command. To run a shell snippet (pipes, redirects, &&), pass it to a shell explicitly: cinna exec bash -c '…'.

With --agent <agent>, run from an account workspace root against the named synced agent (name, slug, or ID) using that child workspace's own token — the agent must already be attached with cinna agent sync.

cinna exec python scripts/main.py
cinna exec pip install pandas
cinna exec bash -c 'ls -la'
cinna exec python -c 'import sys; print(sys.argv)' "a b"
cinna exec --agent crm-agent python scripts/main.py   # from the account root

cinna status

One-shot summary of the agent + current sync state. Includes a backend probe (GET /sync-runtime) that reports whether the stored CLI token is still accepted — valid token, expired token, or no connection. Use cinna set-token to refresh an expired token.

cinna list

List every agent registered on this machine (from ~/.cinna/agents.json). Three columns:

1. Agent — display name on top, full agent ID below.
2. Location — workspace path on top, platform UI link below. Missing directories are flagged in red.
3. Sync — Mutagen session state on top (active / paused / connecting / error), plus a per-agent backend probe (valid token / expired token / no connection) on the bottom. The probes run in parallel with a short timeout so the view stays snappy even with many registered agents.

cinna doctor

Diagnose and repair stale sync state across the whole machine. Over time the per-user registry (~/.cinna/agents.json) and the Mutagen daemon drift out of sync as agents are deleted, environments are spun down, and tokens expire — cinna doctor reconciles the two and heals the leftovers in one pass. Run it from anywhere; it is not workspace-scoped.

It detects and fixes:

• Deleted workspaces — registry entries whose workspace folder (or its .cinna/config.json) is gone. The entry is removed, along with any leftover Mutagen session.
• Halted sessions — sessions stopped on halted-on-root-deletion (the local workspace/ root was deleted) while the agent dir is otherwise intact. Terminated; cinna dev recreates a clean one.
• Dead-remote sessions — sessions stuck retrying a remote env that no longer exists (connecting-beta / beta polling error). Mutagen has no "give up after N failures" option — a session retries forever until paused or terminated — so doctor is the cleanup path for these. Terminated.
• Orphaned sessions — cinna-* sessions with no registry entry at all. Terminated.
• Expired tokens — for account-managed workspaces (those under an account root), the CLI token is re-minted automatically through the parent account token, no pasting required. Standalone workspaces (set up via cinna setup) can only be refreshed with a pasted setup token, so they are reported with a cinna set-token hint rather than changed.
• Expired account token — a sub-agent token can only be re-minted while the account token that mints it is still valid. When the account token has itself expired, doctor probes it once and surfaces a single "renew the account token — run cinna login" finding (listing the blocked sub-agents) instead of a pile of re-mints that would all fail with 401. Run cinna login, then re-run cinna doctor to re-mint the dependents.

cinna doctor              # diagnose, then apply all fixes behind one confirmation
cinna doctor --dry-run    # report problems only; change nothing
cinna doctor --yes        # apply every fix non-interactively

The diagnosis is split into two tables: Will fix (everything doctor can repair — deleted-workspace cleanup, halted/dead/orphaned session termination, account token re-mints) and No automatic fix — manual action needed (standalone expired tokens, which need a pasted setup token and are never touched). Everything actionable is applied together behind a single Apply N fix(es)? confirmation, so the count always matches the "Will fix" table.

cinna disconnect

Stop sync, remove .cinna/ config and generated files (CLAUDE.md, BUILDING_AGENT.md, .mcp.json, opencode.json, mutagen.yml). Workspace files are preserved.

cinna disconnect-all

Scan the current directory for every cinna workspace (directories containing .cinna/config.json), stop each sync session, and delete the directories entirely. Prompts for confirmation and prints a summary of what was removed.

cinna completion [SHELL] [--install]

Output or install shell completion for bash, zsh, or fish.

Workspace Structure

After setup, the agent directory looks like:

my-agent/
  .cinna/                 # CLI config (do not edit)
    config.json
  workspace/              # Continuously synced with the remote env
    scripts/              # Bundle-owned: agent Python scripts
    docs/                 # Bundle-owned: WORKFLOW/ENTRYPOINT/REFINER prompts
    webapp/               # Bundle-owned: dashboard + data endpoints
    knowledge/            # Bundle-owned: static integration docs
    files/                # Bundle-owned: static publisher-shipped assets
    app-data/             # Per-user persistent — NOT shipped in bundle revisions.
                          # Backed by a platform AppDataVolume keyed by (user_id, bundle_id);
                          # mounted on the platform at /app/workspace/app-data.
                          # Survives apply-update and uninstall/reinstall.
      storage/            #   long-lived runtime output (DBs, reports, derived data)
      uploads/            #   all user-supplied file uploads at runtime
                          #   (chat attachments, task attachments, MCP uploads)
      cache/              #   disposable caches
    credentials/          # Backend-managed; visible read-only on your side
    workspace_requirements.txt
    workspace_system_packages.txt
  mutagen.yml             # Sync rules (customizable)
  CLAUDE.md               # Local dev instructions for AI tools
  BUILDING_AGENT.md       # Building mode prompt pulled from the platform
  .mcp.json               # MCP config for Claude Code
  opencode.json           # MCP config for opencode
  .gitignore              # ignores workspace/credentials/ and workspace/app-data/

Persistence tiers mirror the platform's bundle/install model:

• Bundle-owned folders (scripts/, docs/, webapp/, knowledge/, files/, workspace_requirements.txt, workspace_system_packages.txt) are part of what gets snapshotted when a new bundle revision is published. As the developer/publisher, your edits here become the next shipped revision.

• app-data/ is the per-user persistent runtime volume. On the platform it lives in an AppDataVolume keyed by (user_id, bundle_id) — one volume per user per bundle, bind-mounted into the agent container at /app/workspace/app-data. It is not part of bundle revisions: when you publish, only the bundle-owned folders are snapshotted, and every user who installs your bundle gets their own fresh app-data volume. On the platform side the volume survives apply-update (bundle folders are overwritten, app-data is never touched) and uninstall/reinstall (orphaned, not deleted; reattaches by bundle_id). What you see synced to your local workspace/app-data/ is your own developer install's app-data — useful for inspecting runtime output your scripts produce. It's gitignored by default since it's per-user runtime state, not bundle content.

  Where scripts should put what:

  • storage/ — long-lived runtime state (databases, JSON, CSVs, generated reports). Anything the agent must keep across sessions and bundle updates.
  • uploads/ — every user-supplied file at runtime lands here automatically: chat attachments, task attachments, MCP get_file_upload_url uploads. Read from this folder, don't write to it from scripts.
  • cache/ — disposable caches the scripts may rebuild on demand.

• credentials/ is managed by the backend and only readable on your side.

Working with AI Coding Tools

Setup generates MCP server configs for Claude Code (.mcp.json) and opencode (opencode.json), giving your AI tool a knowledge_query tool that searches the agent's knowledge base.

cd my-agent/
claude        # or: opencode

Sync & Conflict Resolution

cinna sync drives Mutagen in two-way-safe mode with VCS-aware ignores (including the backend-managed credentials/ directory, so it never conflicts on files you're told not to edit). When the same file changes on both sides, Mutagen parks a conflict (it does not pick a winner, and does not write .conflict.* files in this mode) — list them with cinna sync conflicts, then clear them with cinna sync resolve --prefer local (your edits win) or --prefer remote (the container's version wins). For a non-interactive flush, cinna sync push / cinna sync pull settle the session and exit.

Large binary files and build artifacts are ignored by default (see mutagen.yml). Add your own ignores there if needed.

Development

git clone https://github.com/opencinna/cinna-cli.git
cd cinna-cli
uv venv && uv pip install -e ".[dev]"
uv run pytest -v
uv run ruff check src/

License

MIT