meridian-governance-cli

Scriptable command-line client for Meridian Governance Router, over meridian-client. Installs as the mdn command — one short command, same shape as mta or claude.

The PyPI package is meridian-governance-cli; the command it installs is mdn. Not to be confused with the unrelated meridian-cli package also on PyPI.

Closer in spirit to the aws or stripe CLIs than to an app's own CLI: subcommands over an API, scriptable, mostly no interactive loop required — with one exception. chat absorbed Local-AI's reference implementation upstream (meridian-agent-v1.0.py, decided 2026-10-18) so customers get one governed resolve -> dispatch -> audit loop from Meridian itself, interactively or scripted, rather than writing that sequence themselves against the SDK.

Install

curl -fsSL https://dev-meridian.mytrustworthy.ai/install.sh | bash

Detects Python 3.9+, installs via pipx if present, otherwise into an isolated venv under ~/.local/share/mdn with a shim on ~/.local/bin. One command, nothing else to figure out. It also drops a small manifest at ~/.meridian/install-manifest.json recording which path it took, so uninstalling doesn't require remembering:

curl -fsSL https://dev-meridian.mytrustworthy.ai/uninstall.sh | bash

Reads that manifest to remove the right thing (pipx uninstall or the venv it created), offers to revoke your API key on the server first (mdn logout --revoke), and separately offers to delete ~/.meridian (saved credentials, chat session logs) — three confirmations, nothing deleted without asking.

Or plain pip, if you'd rather manage the venv yourself (no manifest, so uninstall.sh won't know about it — just pip uninstall when you're done):

pip install meridian-governance-cli

Get credentials

mdn signup          # new org — prompts for org name, email, password
# or
mdn login           # existing account — authenticates, mints a new key
mdn whoami           # confirm what's saved

Credentials live in ~/.meridian/config.json (mode 600 — api_key is in there in plaintext, same tradeoff mta-cli makes for its own token). Override the location with MERIDIAN_CLI_HOME if you want more than one profile.

Commands

mdn resolve my_agent [--project-id ID] [--json]
mdn explain ROUTING_LOG_ID
mdn audit ROUTING_LOG_ID --input-tokens N --output-tokens N

mdn projects create --name NAME [--quality Q] [--cost C] [--risk R]
mdn projects list

mdn agents create PROJECT_ID AGENT_NAME [--pinned-model M] [--quality-override Q] [--allowed-hubs h1,h2]
mdn agents list PROJECT_ID

mdn hubs list
mdn hubs discover

mdn keys list [--json]
mdn keys create [--name NAME]
mdn keys revoke KEY_ID

mdn catalog
mdn logout [--revoke]

mdn chat [--agent NAME] [--code-agent NAME] [--project-id ID] [--max-tokens N]
mdn chat --one-shot "message" [--json] [--max-tokens N]
mdn chat --daemon [--host HOST] [--port PORT] [--refresh-interval SECS]

Every data-returning command without a dedicated human-readable format (explain, audit, catalog, projects list, agents list) prints raw JSON — pipe into jq.

chat

Per-turn flow: resolve(agent) -> catalog/hubs join -> dispatch to the chosen hub's OpenAI-compatible endpoint -> audit. Routing, scoring and governance all live in Meridian; chat just drives the loop. OpenAI-compat hubs only for now (vLLM, Ollama, OpenAI-shaped endpoints) — bedrock/ anthropic/etc. need a protocol translator, not yet added.

• No args: interactive REPL. /help inside it for slash commands (/catalog, /hubs, /refresh, /history, /export, /window, …).
• Three REPL modes, switched with /chat, /code, /manage (default chat):
  • chat — plain conversation, resolve(--agent) (default agent name chat).
  • code — same loop, different binding: resolve(--code-agent) (default agent name code) plus a one-time system preamble. Its own history, separate from chat mode's. This is not a tool-using coding agent — Meridian's dispatch layer never sends tools to the hub, so this is "a different agent/model to resolve against," nothing more. A real tool-execution loop would be new dispatch-layer work, not a REPL mode.
  • manage — no LLM calls at all: /projects, /agents, /keys (list/create, /keys revoke ID) over the account you're logged into. Same functions mdn projects/mdn agents/mdn keys call outside the REPL (meridian_cli/manage.py) — one implementation, not two that can drift.
  • /project [ID] works in any mode: with no argument, shows the active project and lists your org's projects; with an ID, switches to it (validated against your org's actual projects first). Affects the next resolve() call in chat/code mode immediately.
• --one-shot "message": single turn, prints the response, exits. --json emits structured output (response, model, hub, routing_log_id, latency_ms, usage) instead of plain text.
• --daemon: long-lived local OpenAI-compatible endpoint (POST /v1/chat/completions, GET /v1/models, GET /health) — point an existing OpenAI-SDK-based app at http://127.0.0.1:8700 and it works unchanged, with Meridian's governance and audit trail in the middle. The F9 "broker" shape from Local-AI's feedback. Stateless across requests, same as the real OpenAI API: your app holds history and sends the full messages array each call.
• --max-tokens: reasoning models can need a much larger budget than the default (8192) just to reach visible content — confirmed live against a real deployment where the same five-word prompt burned anywhere from ~2400 to the full 8192-token budget purely on reasoning, before any answer. If a response comes back as a token-budget notice instead of text, raise this — unless the prompt matches the pattern below, where raising it won't help.
• Exact-count / constraint-satisfaction prompts can make reasoning never converge, at any budget. Isolated live against qwen3.6:35b: "Say hello in exactly five words." burned the entire reasoning budget at 8192, then again at 16384 — raising --max-tokens didn't make it converge faster, it just filled whatever ceiling it was given. The same model answers the unconstrained version ("Say hello.") normally in 186 reasoning tokens, and a substantive non-counting prompt normally in ~700. So this isn't the model's reasoning budget being too small — it's a specific model failure mode where exact-count constraints seem to trap it in verifying the count. Not a Meridian or CLI bug; a model characteristic to know about before raising --max-tokens in response to a stuck "exactly N words/items" prompt — raising it will not fix that case.

hubs discover

Probes localhost for a local Ollama (:11434/api/tags) or vLLM (:8000/v1/models), proposes registering what it finds (models included), and asks for confirmation before actually registering anything — it never writes to your org's catalog without that confirmation. If nothing responds on localhost, it falls back to prompting for a provider + URL manually, probes that instead, and asks for confirmation the same way. Same wizard from mdn hubs discover and /discover inside mdn chat's manage mode (meridian_cli/manage.py + chat.py's cmd_manage_discover) — one implementation either way.

Deliberately only probes localhost, never a LAN subnet: unlike a single-household tool that can hardcode its own known boxes' IPs, Meridian is multi-tenant — there's no fixed list of "this customer's boxes" to guess at. localhost is the one address that means the same thing for every caller; anything else goes through the manual-entry fallback instead of a scan.

Also why this is a client-side probe at all, not a call to the backend's own /discover endpoint: on a SaaS instance, "localhost" from the server's perspective means the server's own container, not your laptop. Discovery and model-listing both happen in the CLI process so they can actually reach localhost on the machine mdn is running on; the resulting model list is registered directly via register_models, never by asking the server to re-discover something it structurally cannot reach.

What this deliberately doesn't do

• hubs discover only knows Ollama and vLLM, and only probes localhost. No cloud-hub registration commands (bedrock/anthropic/openrouter/etc.) — those need credentials entered deliberately, not something to probe for. Console-only for now.