@llamaventures/cli

Llama Ventures team CLI + MCP server for command.llamaventures.vc.

Public source for low-friction install. Not an open-source product — requires a Llama Ventures team account to do anything useful. See Authenticate.

What this ships

This npm package contains two binaries:

• llama — interactive CLI (humans + bash scripts). Zero deps, native fetch.
• llama-mcp — stdio MCP server (Claude Code, Claude Desktop, Cursor, OpenClaw, Codex, any MCP-native agent). Single dep: @modelcontextprotocol/sdk.

Both share the same auth chain and HTTP client, so behaviour stays in lockstep.

Install

npm i -g @llamaventures/cli

Requires Node 18+ (uses native fetch).

Authenticate

The CLI tries credentials in this order on every call:

1. gcloud auth print-identity-token → Authorization: Bearer … (zero config; recommended for team members)
2. $LLAMA_TOKEN env var → X-Llama-Token (preferred for CI / sandboxed agents)
3. ~/.llama/token (single line, mode 0600) → X-Llama-Token
4. ~/.llama-command/config.json — legacy from CLI v0.1; auto-migrates to ~/.llama/token on first read

If both Bearer and X-Llama-Token are present, both are sent. The server tries Bearer first; on verification failure it falls through to X-Llama-Token.

Zero-config (recommended)

gcloud auth login          # one-time, pick your @llamaventures.vc account
llama auth status          # confirm — should show role + email
llama deal search acme-ai  # ready to go

Manual token

For machines without gcloud, or for stable CI / agent setups:

1. Sign in to https://command.llamaventures.vc

2. Visit /settings/tokens → click Mint Token

3. Save the llc_… value to ~/.llama/token:

   llama token set llc_paste_token_here
   # writes ~/.llama/token (mode 0600), round-trips against /api/me before saving

   Or set it as an env var (preferred for CI):

   export LLAMA_TOKEN=llc_paste_token_here

A team member without an account: ask gavin@llamaventures.vc to mint one for you (he can mint for any email; it auto-creates an inactive user row that he then activates).

External / founder use (no Llama token required)

If you don't have a Llama Command token — you're a founder pitching us, an EA, a prospective hire, or just exploring — the CLI ships a separate pitch command family that talks to our public intake agent at command.llamaventures.vc/external-agent. Same conversation, structured intake, same 12-dimension verdict as the web flow — but driven from your terminal (or your AI agent over MCP).

# Bootstrap a session
llama pitch start --name "Jane Doe" --email "jane@acme.ai"

# Send a message (single-shot, prints reply)
llama pitch say "We're building an AI dev tool for X..."

# Attach your deck / one-pager
llama pitch upload ./deck.pdf

# Or open an interactive REPL
llama pitch

Caps (server-enforced — same as the web flow): 5 sessions/IP/day, 3 sessions/email/day, 30min idle timeout, 100 messages/session, 1M tokens/session.

The MCP server exposes the same surface as pitch_* tools (pitch_start, pitch_send_message, pitch_upload_file, pitch_status, pitch_finalize) — so the founder's own AI agent (Claude / Cursor / OpenClaw / etc.) can help them pitch via Llama's intake agent. True A2A.

CLI command reference

# Auth diagnostics
llama auth status

# Token management
llama token set <llc_...> [--base https://command.llamaventures.vc]
llama token show

# Deals — read
llama deal search <query> [--founder ...] [--owner ...] [--status ...] [--stage ...] [--limit N]
llama deal list [--owner ...] [--status ...]
llama deal show <dealId>

# Deals — write
llama deal create "Company" --description "..." [--source ...] [--stage ...] [...]
llama deal update <dealId> <field> <value>
llama deal delete <dealId>                  # soft-delete (audit-logged)
llama deal restore <dealId>
llama deal trash                            # list soft-deleted

# Ownership
llama claim <dealId>                        # propose self
llama nominate <dealId> --user <userId>
llama nominations list
llama nominations decide <approvalId> accepted|declined
llama approvals list                        # partner queue
llama approvals decide <approvalId> approved|rejected [--note "..."]

# Timeline + posts
llama timeline <dealId>
llama post <dealId> "message" [--link url] [--link-name "name"]

# Brief blocks (ordered, typed: text | link | embed | callout)
llama brief blocks <dealId>
llama brief add-text    <dealId> --heading "..." --body "..."
llama brief add-link    <dealId> --url "..." --label "..." [--description "..."]
llama brief add-embed   <dealId> --url "..." [--label "..."]
llama brief add-callout <dealId> --tone insight|info|warning|success --heading "..." --body "..."
llama brief edit        <dealId> <blockId> [--heading ...] [--body ...] [--url ...] [--label ...] [--tone ...]
llama brief delete      <dealId> <blockId>            # soft
llama brief restore     <dealId> <blockId>
llama brief history     <dealId> <blockId> [--limit 50]
llama brief restore-version <dealId> <blockId> <historyId>

# Collaborators
llama deal collab list    <dealId>
llama deal collab add     <dealId> --user <userId|email>
llama deal collab remove  <dealId> --user <userId|email>
llama deal collab restore <dealId> --user <userId|email>

# Links (URLs attached to a deal — Netlify demos, Gamma decks, etc.)
llama deal link list    <dealId> [--include-deleted]
llama deal link add     <dealId> --url <url> [--label "..."]
llama deal link delete  <dealId> <linkId>
llama deal link restore <dealId> <linkId>

# Brief / persona refresh
llama deal refresh-brief   <dealId> [--force]
llama deal refresh-persona <dealId> <persona>

# Deal facts (AI / human-asserted, with verification)
llama deal fact list   <dealId>
llama deal fact add    <dealId> --category <cat> --claim "<text>" [--source <url>] [--confidence high|medium|low]
llama deal fact verify <dealId> <factId> --status confirmed|disputed [--corrected-value "..."]

# Mentions / inbox
llama mentions                              # my unresolved cues (default)
llama mentions list [--everyone] [--all]
llama mentions show <mentionId>
llama mentions resolve <mentionId>
llama mentions unread                       # badge count

# Skill corrections (persona-owner workflow)
llama skill-correction list <skill-slug> [--include-deleted]
llama skill-correction add <skill-slug> "<rule>" [--deal <uuid>] [--block <blockId>]
llama skill-correction delete <id>

# Wiki (knowledge base)
llama wiki search <query>
llama wiki read <slug>
llama wiki save <slug> --title "..." --content "..." [--sources "url1;url2"]

# Admin event feeds (system admin only)
llama admin auth-events  [--kind X] [--actor email] [--since 24h] [--limit 100]
llama admin deal-events  [--kind X] [--deal <uuid>] [--since 24h]
llama admin agent-events [--kind tool_call|loop_stalled] [--errors-only]

Soft-delete

All deletes through this CLI are non-destructive: brief blocks, collaborators, deal links, and deals themselves use deleted_at markers. Default reads filter trashed rows out; pass --include-deleted on deal link list (or visit /admin for the broader trash view) to see them. Every removal and restore writes an audit-log entry.

MCP server (llama-mcp)

The same package ships a stdio MCP server with 16 tools mirroring the most-used CLI surface. Auth is identical — gcloud → $LLAMA_TOKEN → ~/.llama/token. The server reuses lib/client.mjs so the CLI and MCP can never drift on transport or auth.

Tools registered:

auth_status              deal_search   deal_show
deal_create              deal_update
brief_blocks             brief_add_text  brief_add_link  brief_add_callout
wiki_search              wiki_save
timeline                 post
mentions_list
llama_api                # escape hatch — raw HTTP for endpoints not yet wrapped

llama_api is a generic passthrough modeled on the GitHub MCP server pattern: agents discover it via tools/list and use it for any endpoint not yet given a typed wrapper. Path must start with /api/.

Wire into Claude Desktop / Claude Code

~/.config/claude-desktop/claude_desktop_config.json (macOS: ~/Library/Application Support/Claude/claude_desktop_config.json):

{
  "mcpServers": {
    "llama": {
      "command": "llama-mcp"
    }
  }
}

Restart Claude Desktop. The 16 tools appear under the 🛠️ menu.

Wire into Cursor

~/.cursor/mcp.json:

{
  "mcpServers": {
    "llama": {
      "command": "llama-mcp"
    }
  }
}

Wire into a Codex / OpenClaw / arbitrary stdio MCP client

Most clients accept a command + args config. Run which llama-mcp to find the binary path (/usr/local/bin/llama-mcp or ~/.npm-global/bin/llama-mcp) and point the client at it. Any agent that speaks MCP over stdio works.

Error codes (for agents)

CLI errors include a stable prefix so agents can pattern-match and recover:

• Error[NO_AUTH] — no credentials found. Direct the user to gcloud auth login or llama token set.
• Error[UNAUTHORIZED] — server rejected our credentials. Token revoked, expired, or wrong account selected in gcloud.

The MCP server returns the same errors as isError: true content with the same prefix.

Versioning

Semver. Breaking changes to CLI command shape (renamed flags, removed commands) bump major. Adding a tool or flag bumps minor. Bug fixes bump patch.

The CLI prints its version under llama --version. MCP server reports the same version in its serverInfo.

Reporting security issues

Do not file public GitHub issues for security bugs. Email gavin@llamaventures.vc. See SECURITY.md for scope, response SLA, and the supply-chain posture (Trusted Publishers + provenance + zero-deps CLI).

License

MIT.