@llamaventures/cli

The Llama Ventures CLI & MCP server.
One npm install, one auth chain, two interfaces — humans and AI agents talk to command.llamaventures.vc through the same client.

English · 简体中文

Install · Authenticate · CLI · MCP · External pitch · Agent briefing · Security · Changelog

Public source for low-friction install. Not an open-source product. Most operations require a Llama Ventures team account (gavin@llamaventures.vc mints tokens). The one exception is the public pitch family — see External pitch.

What's in the box

@llamaventures/cli
├── bin/llama          interactive CLI for humans + bash
└── bin/llama-mcp      stdio MCP server, 20 tools — for any MCP-native agent

Both binaries share lib/client.mjs — the same auth chain, same HTTP client, same error format. CLI and MCP can never drift on transport or identity. Zero runtime dependencies for the CLI itself; the bundled MCP server depends only on @modelcontextprotocol/sdk (Anthropic-maintained, pinned exact).

flowchart LR
  Human([🧑‍💻 Human])           --> CLI[bin/llama<br/>argv parser]
  Agent([🤖 MCP-native agent]) --> MCP[bin/llama-mcp<br/>stdio JSON-RPC]
  CLI  --> Client[lib/client.mjs<br/>auth · fetch · errors]
  MCP  --> Client
  Client -- HTTPS --> API[(command.llamaventures.vc)]
  classDef src fill:#dcfce7,stroke:#166534,color:#14532d
  classDef edge fill:#dbeafe,stroke:#1e40af,color:#1e3a8a
  class Human,Agent edge
  class CLI,MCP,Client src

Install

npm i -g @llamaventures/cli

Requires Node 18+ (uses native fetch and ESM). CI runs the matrix on 18 / 20 / 22.

Verify:

llama --version
llama auth status     # round-trips against /api/me

The same install puts llama-mcp on your PATH for the MCP server — no second package.

Upgrading from npm link? The CLI used to live in the llama-os/cli/ directory and was distributed via npm link. As of CLI v1.x it ships as @llamaventures/cli. Run npm i -g @llamaventures/cli@latest; the legacy directory keeps working during the soak window but is no longer the source of truth. See llama-os/cli/DEPRECATED.md.

Authenticate

The client tries credentials in this order, on every call:

Both Bearer and X-Llama-Token are sent if both exist. The server tries Bearer first; on verification failure it falls through to X-Llama-Token. Inspect the resolved identity any time with llama auth status.

Zero-config — recommended for team members

gcloud auth login          # one-time; pick your @llamaventures.vc account
llama auth status          # → role + email
llama deal search acme-ai  # ready

Manual token — for machines without gcloud, or stable CI

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

2. Open /settings/tokens → Mint Token.

3. Save the llc_… value:

   llama token set llc_paste_token_here
   #  → writes ~/.llama/token (mode 0600)
   #  → round-trips /api/me before saving — bad token never lands on disk
   

   Or, in CI / one-shot environments:

   export LLAMA_TOKEN=llc_paste_token_here
   

Don't have an account? Email gavin@llamaventures.vc. Any email — including non-@llamaventures.vc — can be granted a token; the system admin mints it via /settings/tokens. Token first-use auto-creates the user row.

CLI tour

The CLI is the canonical interface. The HTTP API beneath it is stable, but the CLI handles auth, error formatting, and forward-compatibility across server schema changes — prefer the CLI even from inside scripts.

# Auth + tokens
llama auth status
llama token set <llc_...>
llama token show

# Pipeline — read
llama deal search "acme ai"
llama deal list --owner alex --status Diligence
llama deal show <dealId>

# Pipeline — write
llama deal create "Acme AI" --description "..." --source Gavin
llama deal update <dealId> status Diligence
llama deal delete  <dealId>     # soft (audit-logged)
llama deal restore <dealId>

# Deal Brief — ordered, typed blocks (text · link · embed · callout)
llama brief blocks       <dealId>
llama brief add-text     <dealId> --heading "..." --body "..."
llama brief add-link     <dealId> --url "..." --label "..."
llama brief add-callout  <dealId> --tone insight --heading "..." --body "..."
llama brief edit         <dealId> <blockId> [--heading ...] [--body ...]
llama brief history      <dealId> <blockId>

# Ownership + approvals
llama claim       <dealId>
llama nominate    <dealId> --user <userId>
llama approvals   list
llama approvals   decide <approvalId> approved --note "..."

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

# Wiki
llama wiki search "<query>"
llama wiki save <slug> --title "..." --content "..."

# Mentions inbox
llama mentions
llama mentions resolve <mentionId>

Run llama --help for the full surface (~40 commands across deals, briefs, ownership, timeline, facts, wiki, mentions, skill corrections, and admin event feeds). Soft-delete is the default everywhere — every removal is reversible and audit-logged via deal_events.

Error codes — for agents

The CLI's stderr exit messages start with stable, parseable prefixes:

The MCP server returns the same prefixes inside isError: true content so agents can pattern-match without parsing prose.

MCP server

The bundled llama-mcp is a stdio Model Context Protocol server exposing 19 typed tools that mirror the most-used CLI surface. Every tool is named and scoped — there is no generic API passthrough, by design (a public-package escape hatch reachable from a prompt-injectable agent context is exactly the shape we want to avoid).

auth_status

deal_search            deal_show
deal_create            deal_update

brief_blocks           brief_add_text
brief_add_link         brief_add_callout

timeline               post

wiki_search            wiki_save

mentions_list

pitch_start            pitch_send_message
pitch_upload_file      pitch_status
pitch_finalize

Auth is identical to the CLI's chain (gcloud → $LLAMA_TOKEN → ~/.llama/token). The agent_briefing MCP prompt also returns AGENT_BRIEFING.md verbatim, so any new agent loading the server can self-onboard without leaving the protocol.

Wire into your agent

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

claude mcp add llama -- llama-mcp

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

If you're new and want the agent to onboard itself, run llama agent-onboard from the CLI or fetch the agent_briefing prompt from the MCP server. It's the workflow contract — autonomy levels, attribution grammar, error recovery, anti-pollution rules.

External pitch — no Llama account required

If you're a founder pitching us, an EA, or a prospective hire without a Llama Command token, the CLI ships a pitch command family (and the parallel pitch_* MCP tools) that talks to our public intake agent at command.llamaventures.vc/external-agent. Same conversation, same structured 12-dimension verdict — driven from your terminal or your own AI agent.

llama pitch start --name "Jane Doe" --email "jane@acme.ai"
llama pitch say "We're building an AI dev tool for X..."
llama pitch upload ./deck.pdf
llama pitch                       # interactive REPL

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

This is genuine agent-to-agent: your AI helps you tell the story, our intake agent extracts the structured fields and produces the verdict.

Stability

• Versioning: SemVer. Renaming or removing a CLI command bumps major. Adding a tool, command, or flag bumps minor. Bugfixes bump patch. The CLI prints --version; the MCP server reports the same value in its serverInfo.
• Backwards compatibility: The wire format (Bearer / X-Llama-Token) and the Error[…] prefixes are part of the public contract and won't change inside a major version.
• Server schema drift: When the API gains an endpoint, the CLI / MCP gain a typed wrapper in the next minor release. While you wait, the llama CLI itself ships the full llama command surface (40+ commands) — use it for ad-hoc HTTP work that the MCP doesn't yet wrap.

See CHANGELOG.md for the per-version log.

Security

• @llamaventures/cli is published via npm Trusted Publishers — no NPM_TOKEN lives in repo secrets. Each release ships with --provenance (sigstore-signed); the npm registry shows a Provenance badge traceable to the exact GitHub Action workflow + commit.
• Minimal dependency tree. The CLI is zero-deps. The MCP server depends only on @modelcontextprotocol/sdk, pinned exact.
• Branch protection on main; Dependabot, secret scanning, and push-protection are enabled.
• Tokens: stored locally at ~/.llama/token mode 0600. Server-side they are stored as sha256 hashes — plaintext only ever exists in the user's possession.

Reporting a vulnerability: see SECURITY.md. Do not file public GitHub issues for security bugs.

Contributing

This is an internal tool maintained by Llama Ventures. PRs from team members are welcome — see CONTRIBUTING.md for the local dev loop, release flow (Trusted Publishers + GitHub Releases), and the conventions we follow (zero-deps, lockstep CLI/MCP, stable Error[…] prefixes).

External contributions: feel free to open issues for documentation gaps or broken flows. Feature requests for non-team workflows are best directed at the external pitch path instead.

License

MIT — © 2026 Llama Ventures, Inc.