Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Add Tavily Search and Extract as keyless remote MCP tools for selected NanoClaw agent groups. Use when installing Tavily web search or URL extraction without an API key.
.claude/skills/nanocoai-add-tavily-tool/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-04 | ✗→✓ | ▲ Improved | 57% | 0% |
| case-07 | ✗→✓ | ▲ Improved | 66% | 0% |
| case-08 | ✗→✓ | ▲ Improved | 26% | 0% |
| case-09 | ✗→✓ | ▲ Improved | 35% | 0% |
| case-11 | ✗→✓ | ▲ Improved | 20% | 0% |
Install the pinned mcp-remote bridge in the agent image and register Tavily's remote MCP server for each selected agent group. The MCP server supplies its tool descriptions and input schemas at runtime.
The registered server exposes:
mcp__tavily__tavily_searchmcp__tavily__tavily_extractThe registration is provider-agnostic: any provider with MCP support picks it up (Claude, OpenCode, and Codex all do). Groups on the Claude provider already have the built-in WebSearch and WebFetch tools (container/agent-runner/src/providers/claude.ts), so the skill adds the most for groups on other providers, and for Tavily's structured extraction anywhere.
Check whether the bridge is already in the image manifest, then list the groups:
bashgrep -n '"mcp-remote"' container/cli-tools.json || true ncl groups list
Ask which agent groups should receive Tavily. If mcp-remote is already present at a pinned version, reuse the existing entry instead of adding a second one.
Add this object to the top-level array in container/cli-tools.json when an entry named mcp-remote is not already present:
json{ "name": "mcp-remote", "version": "0.1.38" }
Keep the JSON valid and limit the entry to the two fields shown; this package does not require a native build-script opt-in.
Copy the dependency guard into the host test tree:
bashcp .claude/skills/add-tavily-tool/tavily-manifest.test.ts src/tavily-manifest.test.ts
Build the image and run the guard:
bash./container/build.sh pnpm exec vitest run src/tavily-manifest.test.ts
The manifest is the only source-backed integration point. Per-group MCP registration is runtime state stored through ncl, so it has no in-tree line for a registration test to guard.
config add-mcp-server and groups restart are approval-gated. Run from inside a container they return approval-pending immediately; that is not an error. Wait for the admin's approval and the follow-up system message before moving on to Phase 4.
For each selected <group-id>, register one server named tavily:
bashncl groups config add-mcp-server \ --id <group-id> \ --name tavily \ --command mcp-remote \ --args '["https://mcp.tavily.com/mcp/","--transport","http-only","--enable-proxy","--header","X-Tavily-Access-Mode:keyless","--header","X-Client-Name:nanoclaw","--ignore-tool","tavily_crawl","--ignore-tool","tavily_map","--ignore-tool","tavily_research"]' \ --env '{}'
The keyless header enables Tavily's IP-based allowance. The client-name header attributes calls to NanoClaw. The tool filters leave only Search and Extract available.
Restart each selected group:
bashncl groups restart \ --id <group-id> \ --message "Tavily Search and Extract are installed. Run one Tavily search with max_results 1 and report whether it succeeds."
Confirm the stored configuration contains one tavily server with both headers:
bashncl groups config get --id <group-id>
Then check the selected agent's test response. The call must use mcp__tavily__tavily_search. Tavily Crawl, Map, and Research must not appear in the Tavily namespace.
The keyless allowance is shared by every group on the host, so it can run out. Install standing instructions so the agent offers the paid-key upgrade at that moment instead of dead-ending. For each selected group:
bash docker inspect onecli --format '{{range .Config.Env}}{{println .}}{{end}}' | grep '^APP_URL='
If the value is a loopback or container-bridge address (127.0.0.1, 172.17.0.1, host.docker.internal), ask the operator which URL they open the OneCLI dashboard at, suggesting http://127.0.0.1:10254 as the default. A public or tailnet APP_URL needs no question.
curl -fs <dashboard-url>/connections/custom must returnHTTP 200. If it does not (older OneCLI without the prefill route), replace step 2 of the template with: "Ask an operator to run, on the host: onecli secrets create --name tavily --type generic --host-pattern mcp.tavily.com --header-name Authorization --value-format 'Bearer {value}' --file <key-file>".
{{ONECLI_DASHBOARD_URL}} inupgrade-instructions.md with the resolved URL and write the block into groups/<group-folder>/instructions.prepend.md: replace an existing <!-- tavily-upgrade:start --> to <!-- tavily-upgrade:end --> block in place, append otherwise. Do not write into groups/<group-folder>/CLAUDE.md; it is regenerated at spawn and appended blocks are lost.
dialog loads with host mcp.tavily.com prefilled. If they supplied a public URL while APP_URL was a loopback address, suggest setting the public URL in the OneCLI dashboard (Settings, Instance) so future links stay stable.
ncl groups restart --id <group-id>.If Tavily returns HTTP 429 or monthly_cap_reached_bonus_eligible, the keyless allowance is exhausted. With Phase 5 installed the agent offers the upgrade on its own: the user creates a free API key and stores it through the prefilled dashboard link; the key lands in the OneCLI vault and the gateway injects it into the bridge's requests. The agent then re-registers the server without the X-Tavily-Access-Mode:keyless header and restarts the group. The agent never sees the key.
command not found: mcp-remote: rebuild the image, then restart the group.tavily MCP entry, thenrestart it.
--ignore-tool pairs.429 or monthly_cap_reached_bonus_eligible: the keyless allowance isexhausted; see Keyless limit for the OneCLI upgrade path.
groups/<group-folder>/instructions.prepend.md contains the tavily-upgrade block (Phase 5) and restart the group. A session that already discussed the limit keeps reasoning from that history; /clear starts a clean one.
See REMOVE.md for the idempotent removal procedure.
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→fail | 16,128 | 16,114 | -0% | 1 | 1 | 0% | 260 | 2,171 | +735% | 0 | 0 | — |
case-06 | pass→pass | 14,983 | 12,061 | -20% | 1 | 1 | 0% | 1,564 | 3,000 | +92% | 0 | 0 | — |
case-02 | fail→fail | 16,668 | 19,716 | +18% | 1 | 1 | 0% | 263 | 2,723 | +935% | 0 | 0 | — |
case-03 | fail→fail | 14,192 | 15,134 | +7% | 1 | 1 | 0% | 262 | 2,162 | +725% | 0 | 0 | — |
case-04 | fail→pass | 13,877 | 18,199 | +31% | 1 | 1 | 0% | 1,685 | 2,648 | +57% | 0 | 0 | — |
case-05 | fail→fail | 17,757 | 16,408 | -8% | 1 | 1 | 0% | 1,915 | 2,288 | +19% | 0 | 0 | — |
case-07 | fail→pass | 14,263 | 8,680 | -39% | 1 | 1 | 0% | 1,521 | 2,521 | +66% | 0 | 0 | — |
case-08 | fail→pass | 15,923 | 7,290 | -54% | 1 | 1 | 0% | 1,841 | 2,324 | +26% | 0 | 0 | — |
case-09 | fail→pass | 16,478 | 7,949 | -52% | 1 | 1 | 0% | 1,781 | 2,409 | +35% | 0 | 0 | — |
case-10 | fail→fail | 15,895 | 9,689 | -39% | 1 | 1 | 0% | 1,739 | 2,529 | +45% | 0 | 0 | — |
case-11 | fail→pass | 19,416 | 10,246 | -47% | 1 | 1 | 0% | 2,463 | 2,944 | +20% | 0 | 0 | — |
case-12 | fail→pass | 15,747 | 6,861 | -56% | 1 | 1 | 0% | 1,732 | 2,230 | +29% | 0 | 0 | — |
case-13 | fail→pass | 10,547 | 7,308 | -31% | 1 | 1 | 0% | 852 | 2,233 | +162% | 0 | 0 | — |
case-14 | pass→pass | 13,653 | 7,919 | -42% | 1 | 1 | 0% | 1,409 | 2,489 | +77% | 0 | 0 | — |
case-15 | pass→pass | 15,906 | 9,391 | -41% | 1 | 1 | 0% | 1,825 | 2,689 | +47% | 0 | 0 | — |
case-21 | fail→fail | 11,426 | 6,517 | -43% | 1 | 1 | 0% | 1,008 | 2,108 | +109% | 0 | 0 | — |
case-16 | fail→fail | 14,729 | 9,179 | -38% | 1 | 1 | 0% | 1,624 | 2,570 | +58% | 0 | 0 | — |
case-17 | fail→pass | 14,779 | 7,714 | -48% | 1 | 1 | 0% | 1,605 | 2,346 | +46% | 0 | 0 | — |
case-18 | fail→pass | 19,114 | 9,917 | -48% | 1 | 1 | 0% | 2,515 | 2,884 | +15% | 0 | 0 | — |
case-19 | pass→pass | 15,757 | 7,241 | -54% | 1 | 1 | 0% | 1,729 | 2,217 | +28% | 0 | 0 | — |
case-20 | pass→pass | 13,185 | 7,972 | -40% | 1 | 1 | 0% | 1,191 | 2,418 | +103% | 0 | 0 | — |
case-22 | pass→pass | 11,206 | 6,848 | -39% | 1 | 1 | 0% | 1,111 | 2,225 | +100% | 0 | 0 | — |
case-23 | fail→fail | 12,218 | 8,864 | -27% | 1 | 1 | 0% | 1,184 | 2,553 | +116% | 0 | 0 | — |
case-24 | fail→fail | 7,551 | 7,123 | -6% | 1 | 1 | 0% | 415 | 2,240 | +440% | 0 | 0 | — |
DecimalAI ran this skill against gemini-3.6-flash twice over the same eval suite — once with the skill loaded and once without — and compared the two runs case by case. 24 cases were attempted, and 20 counted toward the lift figure. The other 4 produced results that are not comparable between the two arms, so they are excluded from the headline rather than averaged into it. The headline lift of +38 percentage points is the difference between those two pass rates over the 20 comparable cases.
Without the skill loaded, the model failed this case. With it loaded, the same prompt on the same model passed. This is one improved case from the latest verified run; every case, including any that regressed, is in the table above.
Other measured skills in the registry, with their headline benchmark lift.