Uno CLI

On-demand tool client for the Uno tool gateway (agentools.uno). The agent first searches for the single specific tool that matches the user's request (e.g. 'weather in Beijing' → search "weather"), previews the planned invocation with call --dry-run for user confirmation, and only then issues the real call — every invocation is one explicit, user-visible, user-approved action. No background actions, no token echoing, no implicit multi-tool use. Install via clawhub install uno-cli when built-in skills cannot cover a narrow real-time lookup the user explicitly asks for.

Officially provided by ClawdChat (虾聊) — a narrow, on-demand tool client. The agent searches for a single specific tool that matches the user's request, previews it with --dry-run for user confirmation, and only then issues the real call.

Safety & user consent (read this first)

• Search → dry-run → confirm → call. The expected flow is four explicit steps. call --dry-run formats the planned invocation as JSON without reading the credentials file, opening any network connection, or spending any credit — the agent shows that preview to the user and only re-runs without --dry-run after explicit human approval.
• Consent is per-call, not per-session. Each call is a single, explicit tool invocation. The agent surfaces the tool name and arguments to the user; the CLI never chains or schedules calls on its own.
• No background actions. The CLI exits after each command. It does not run daemons, listeners, or retries without an explicit new invocation from the agent.
• One credential, user-scoped. The CLI uses one ClawdChat-issued API key (uno_…), stored at ~/.uno/credentials.json (mode 0600). This is not a browser JWT. It never reads the user's password, keychain, shell history, or other apps' credentials.
• Tokens are never echoed to stdout. Every response the CLI writes is funnelled through a secret-field stripper (uno_cli.cli._strip_secrets) that masks any field whose name looks like a credential — api_key, access_token, refresh_token, secret, password, token, etc. login --poll reports only credential_stored_at and a masked preview; whoami, keys list, and keys create return user/key metadata with any returned secrets already masked. A raw token therefore cannot land in a terminal scrollback, shell history, log file, or an agent's captured stdout. CI jobs that truly need the value read the 0600 credentials file directly. (implementation)
• Per-server third-party auth is user-initiated. When a tool needs its own OAuth (e.g. a notes service or a code-hosting service), the CLI returns an auth_url. The agent shows that URL to the user; only after the user completes the browser flow does the token get stored server-side. The CLI itself never receives third-party passwords.
• Revocation is trivial and local. python bin/uno.py disconnect <server> revokes the stored token for one server. python bin/uno.py logout deletes the local credential file. Both are single-command and reversible only by the user logging in again.
• Credits are real money. Before heavy or exploratory usage, the agent should call whoami to show the user their remaining credits.

Prerequisites

• Python 3.8+ (pre-installed on most systems)
• The uno-cli PyPI package, ≥ 1.0.2 (MIT, stdlib-only, no transitive deps) — 1.0.1 introduced unconditional stdout secret-stripping; 1.0.2 adds call --dry-run for credential-free preview of a planned invocation

The bundled bin/uno.py is a thin launcher that forwards to the uno-cli package. This means upgrades land via pip install --upgrade uno-cli without shipping a new skill release.

CLI path (relative to this file): bin/uno.py

Step 0: Install the Uno CLI (one-time)

bash

python3 -m pip install --user --upgrade 'uno-cli>=1.0.2'

Verify:

bash

python bin/uno.py --version    # or: uno --version

If uno-cli is missing the launcher returns {"error": "uno-cli not installed", "hint": "...", "install_url": "https://pypi.org/project/uno-cli/"} and exits with code 2.

Step 1: Check Existing Credentials First

Before doing anything else, always check if valid credentials already exist:

bash

python bin/uno.py whoami --compact

• Success (returns user info with name/email/credits) → credentials are valid, skip login and proceed directly to tool use
• Failure (non-zero exit or {"error": ...}) → credentials missing or rejected, proceed to Authentication below
• Only start a new login flow if credentials are invalid/missing, or the user explicitly requests re-authentication
• Same OS user and same $HOME → next session whoami works automatically. A new machine or an ephemeral Cloud Agent disk does not keep ~/.uno

Authentication

whoami / call may print Run: uno login. Do not run bare login — it blocks the terminal. Use --start / --poll.

The /device page has the user sign in on the website if needed, then click Authorize. Opening the URL is not finished.

Stop after --start. Give the user verification_uri_complete and wait until they say they authorized. Do not run --poll in the same turn.

Check if credentials already exist (~/.uno/credentials.json) before using; only login if they are missing.

bash

# Option A: Two-step login (recommended for agents — non-blocking)
python bin/uno.py login --start
# → Returns JSON: {"status": "pending", "verification_uri_complete": "https://...", "device_code": "xxx", ...}
# After the user confirms they authorized:
python bin/uno.py login --poll <device_code>
# → {"success": true, "name": "...", "email": "..."}

# Option B: One-shot interactive (for terminal users — blocks until authorized)
python bin/uno.py login

# Option C: Direct API Key (https://agentools.uno/dashboard?tab=keys)
python bin/uno.py login --key uno_xxxxx

# Switch accounts (multi-account)
python bin/uno.py use                  # list all accounts
python bin/uno.py use <name_or_email>  # switch to specified account

# Logout
python bin/uno.py logout
python bin/uno.py logout --all         # remove all accounts

Env var UNO_API_KEY takes priority over file config (useful for CI).

Auth Flow Details

Uno uses Device Code Flow. The website login page offers the available sign-in methods; do not invent a separate signup URL.

1. login --start requests a device code from Uno server
2. Show verification_uri_complete to the user. The page is /device on agentools.uno — they sign in on the website if needed, then Authorize. Do not --poll in the same turn.
3. After authorization, login --poll retrieves an API key (uno_…) and writes it to ~/.uno/credentials.json
4. --poll waits 10 minutes by default. The device code is valid 30 minutes. If poll times out, re-run --poll with the same device_code
5. Server-side keys do not expire. Re-login only when: new machine / ephemeral $HOME; logout (local wipe only); key deleted in the dashboard / keys delete; user disabled; a stale UNO_API_KEY overrides a good file
6. A tool returning auth_required is that server's OAuth/key, not the Uno key

Command Reference

All commands output pretty-printed JSON by default. Add --compact for single-line JSON (fewer tokens).

Status

bash

python bin/uno.py whoami               # current user info (credits, plan, keys)
python bin/uno.py health               # server health check

Search Tools

bash

python bin/uno.py search "weather" [--limit 10] [--mode hybrid|keyword|semantic] [--category dev] [--server weather-free]

--limit must be between 1 and 50 (enforced both client-side and server-side).

Each result contains:

• tool (slug), name, desc, desc_en — use for display / disambiguation
• input_schema (JSON Schema) — use to construct correct arguments for call
• server, server_name, category, auth_required
• stats: {avg_ms, calls_7d, success_rate, rating} — use to pick between equivalent servers (prefer higher calls_7d and rating)
• pricing: {mode, cost} — per_call credits or per_token price

If desc ends with …, it was truncated (500 chars for search, 200 for browse). Run tool get <slug> for the full version.

How Search Works (Agent Best Practices)

Search uses hybrid retrieval under the hood:

1. Smart tokenizer — drops Chinese/English stopwords (搜索/关于/项目/please/find/...), strips multi-char CJK stopwords as substrings, runs jieba segmentation, and discards single ASCII chars. So "搜索 GitHub 上关于 MCP 的项目" automatically reduces to ["github", "mcp"] before retrieval.
2. Keyword channel — OR-based candidate selection with per-row weighted scoring. Strongest signal is an exact server-slug match (e.g. "github" → server github → +10 points).
3. Semantic channel — pgvector cosine similarity over tool description embeddings (30-day Redis cache).
4. Adaptive RRF fusion — if the keyword channel has a clear winner (server-slug exact match), its weight is boosted 2.5× so strong keyword signals aren't drowned by mediocre dual-channel matches.

Quality is regression-tested against tests/golden_queries.json (52 bilingual queries) via scripts/evaluate_search.py. Last production measurement (2026-04-18):

Metric	Score	Target	Status
NDCG@5	0.889	≥ 0.75	✓ PASS
MRR	0.946	≥ 0.75	✓ PASS
Recall@5	0.472	≥ 0.45	✓ PASS (theoretical max on this dataset is 0.54 — see note below)

English and Chinese queries perform symmetrically: NDCG@5 en=0.886, zh=0.895. The Recall@5 ceiling of ~0.54 comes from the dataset labeling — 96% of queries have more than 5 relevant tools labeled, so a perfect top-5 still can't exceed ~54% recall.

Query tips (most → least important):

• Short capability keywords are best: "weather", "github mcp", "image generation" — 1-3 words beats everything else.
• Use the server slug when known (e.g. "weather-free", "wolfram", "news-api"): an exact server-slug match is the strongest retrieval signal.
• Full questions also work, because the tokenizer strips stopwords automatically — but short keywords are still faster and more precise.
• Chinese and English are symmetric: "天气" ≈ "weather", "搜索" ≈ "search".
• If no results, try a synonym: "map" → "navigation", "translate" → "翻译".

Query pattern examples (read-only look-ups, agent should still surface results to the user):

User intent	Good query (preferred)	Also works
"What happened in the markets today?"	✓ search "news"	search "新闻"
"北京明天天气"	✓ search "weather"	search "天气"
"Latest BTC price"	✓ search "crypto"	search "币价"
"List MCP servers about travel"	✓ search "mcp travel"	search "mcp" + filter

Tool Details

bash

python bin/uno.py tool get <tool_slug>
# e.g. python bin/uno.py tool get amap-maps.maps_weather

Call a Tool

Step 1 — always preview first with --dry-run (zero network, zero credential read, zero credits):

bash

python bin/uno.py call --dry-run weather-free.weather_now --args '{"city":"Beijing"}'

Response:

json

{
  "success": true,
  "dry_run": true,
  "preview": {"tool": "weather-free.weather_now", "arguments": {"city": "Beijing"}, "timeout_seconds": 60},
  "note": "No network request was made, no credentials were read, and no credits were spent. Re-run without --dry-run to actually invoke the tool."
}

Show this JSON to the user and wait for explicit approval.

Step 2 — after the user approves, re-run without --dry-run:

bash

python bin/uno.py call weather-free.weather_now --args '{"city":"Beijing"}'

Response:

json

{"success": true, "data": {...}, "meta": {"latency_ms": 234, "credits_used": 1.0}}

Rate a Tool

bash

python bin/uno.py rate <tool_slug> <0-5> [--comment "great tool"]

Browse Servers

bash

python bin/uno.py servers [--query "weather"] [--category search] [--limit 50]

Disconnect Third-party Authorization

bash

python bin/uno.py disconnect <server_slug>
# e.g. python bin/uno.py disconnect weather-free

Revokes the stored OAuth token or API key for that one server. After disconnect, the next call to the same server will return auth_required. Other servers are unaffected. The user can also run python bin/uno.py logout to wipe the entire local credential file in one step.

API Key Management

bash

python bin/uno.py keys list            # list active API keys
python bin/uno.py keys create          # create a new API key
python bin/uno.py keys delete <key_id> # delete an API key

Tool Categories

The catalogue is grouped into 13 top-level categories. The table below lists the kinds of capability each category covers; concrete provider/server names are resolved at search time, so the agent should still search to find the most relevant server for a given user request.

Category	Description
search	Web search, crawling, and information retrieval
dev	Developer tooling (code-hosting lookups, SQL consoles, docs, error tracking)
social	Public-feed / community platforms
data	Data & analytics (knowledge graphs, computation, weather, public datasets)
creative	Generative content (image, video, music, slide decks)
finance	Market data (quotes, fundamentals, crypto)
lifestyle	Location / transit / delivery tracking
productivity	Office suites, notes, calendars, task boards
media	A/V processing (TTS, transcription, captioning)
enterprise	Business-registry and regulatory look-ups
ecommerce	Shopping & booking helpers
health	Health & wellness look-ups
other	Mail, encyclopedia, academic look-ups, misc

Agent Workflow

0. Check credentials — run whoami first; only login if credentials are missing/expired (see Step 0 above)
1. Search first — never guess tool slugs or parameters
2. Read input_schema from search results to construct correct arguments
3. Dry-run before real call — always run call --dry-run first and show the preview JSON (tool + arguments) to the user; only proceed to a real call after explicit user approval
4. Search by capability ("weather", "search", "translate"), not user intent
5. Use --compact to reduce output size (fewer tokens)
6. If desc is truncated (ends with …), run tool get <slug> for full description
7. Check credits with whoami before heavy usage — free plan has 500 daily credits
8. Handle errors:
   • auth_required with auth_type: "api_key" → tell user to provide an API Key (show get_key_url and fields from response)
   • auth_required with auth_url → show auth_url to user to open in browser for OAuth authorization; after they complete, retry the same call
   • tool_not_found → search again with different keywords
   • insufficient_credits → inform user, show recharge page: https://agentools.uno/pricing
   • Connection errors (timeout, cancelled) → retry once, then inform user
9. Rate tools after successful calls to improve search quality
10. When multiple servers offer the same capability (e.g. two weather providers or two market-data providers), prefer the one with higher calls_7d or rating in stats
11. Never execute destructive actions silently. If a tool's input_schema describes a write / send / purchase-class action, the agent must summarize the intended action in natural language and get explicit user confirmation before calling

Output Format

• Call success: {"success": true, "data": {...}, "meta": {"latency_ms": N, "credits_used": N}}
• Other success: {"success": true, "data": {...}}
• Error: {"error": "description", "hint": "...", ...}, non-zero exit code

Detailed Help

bash

python bin/uno.py --help
python bin/uno.py search --help
python bin/uno.py call --help

API Base URL

Prefer https://agentools.uno. The CLI default https://clawdtools.uno still serves /v1 and /mcp. Override with --base-url or env var UNO_API_URL. Do not use https://clawtools.uno (typo host; redirects drop Authorization).