Skip to content

fix(integrations): correct Nimble page tool list and naming - #1101

Open
thierrypdamiba wants to merge 4 commits into
mainfrom
docs/nimble-accuracy-fixes
Open

fix(integrations): correct Nimble page tool list and naming#1101
thierrypdamiba wants to merge 4 commits into
mainfrom
docs/nimble-accuracy-fixes

Conversation

@thierrypdamiba

@thierrypdamiba thierrypdamiba commented Jul 29, 2026

Copy link
Copy Markdown
Contributor

Follow-up to #1097 — the "improved version" from the Slack thread, since #1097 was intentionally merged as-is to unblock shipping the URL.

Verified against the live server, not just published docs: tools/list handshake against mcp.nimbleway.com/mcp (Nimble 3.4.3, protocol 2025-06-18) using our trial account.

The tool list was wrong: 27 tools, not 6

Nimble's docs describe six capability families, and the page reproduced those as six rows. The live server exposes 27 tools — each family ships multiple tools for synchronous, asynchronous, and status/listing calls.

This matters in Arcade specifically. The gateway tool picker lists individual tools, so someone who reads "6 capabilities" and then sees 27 checkboxes has no way to map between them. The table now carries a per-capability tool count and the total:

Capability Tools
Search 1
Extract 2
Crawl 4
Map 1
Extract templates 7
Web search agents 11
Task results 1

Also added web search agents as a capability, which the original page omitted entirely — it's the largest family on the server at 11 tools.

Async polling gotcha

Many Nimble tools return a task ID and require polling a companion tool (nimble_task_results, nimble_agents_run_status, nimble_crawl_status). Since selecting a subset of tools is the normal gateway workflow, it's easy to build a gateway with only the starting tools and leave your agent kicking off jobs it can't read back. Now called out.

Tool naming was misleading

The page said tools appear as <your-server-id>.<tool>. Nimble prefixes its tools with nimble_, so with server ID nimble the real name is nimble.nimble_search — a doubled prefix that surprises people mid-integration. Confirmed nimble_search exists on the live server.

Cost section: attributed, not asserted

I initially removed this believing it was unsourced. That was my error — I'd only checked docs.nimbleway.com. It's published on nimbleway.com/mcp and in Nimble's MCP blog post:

"Compound accuracy over time through a Proprietary Index and Memory that gets smarter with every query"
"Cut token costs by retrieving exactly what's needed — no redundant searches, no parsing raw pages with an LLM"

Restored with two adjustments:

  1. Attributed to Nimble rather than asserted as fact — product claims about a third party's internals that we can't verify from outside, on a partner-facing page.
  2. Dropped the detail that outran the source. The original said the cache stores "sources, search history, and responses" and the memory layer "extracts key learnings after each run and feeds them into the next search plan." Nimble publishes "Proprietary Index and Memory" and "auditable Search Plans" — not that breakdown.

Kept the practical MCP token caveat alongside it. The two sources genuinely conflict: the product page sells retrieval efficiency, while Nimble's docs warn MCP "can consume tokens quickly" and point high-volume users at their CLI skill. Readers should see both.

Stale tool names on Nimble's marketing page

Worth passing back to Nimble: nimbleway.com/mcp advertises nimble_deep_web_search, nimble_google_maps_search, and nimble_google_maps_reviews. None of these exist on the live server. Their docs were right and their product page is out of date — a heads-up they'd probably want, since it's the page prospects read.

Style

Changelog said "Partner MCP Server"; STYLEGUIDE.md specifies lowercase "MCP server". Page body was already correct.

Verified, not changed

  • Authorization: Bearer ${secret:NIMBLE_API_KEY} matches both Nimble's "store the raw key, Bearer prefix goes in the header" instruction and the ${secret:NAME} syntax in add-remote-servers — confirmed working against the live server
  • All external links and internal anchors resolve, including #configure-advanced-settings
  • pnpm build passes, route compiles, Vale clean (two remaining warnings are pre-existing false positives on UI labels after colons)
  • Nimble absent from search/_meta.tsx, but so is Tavily — partner pages are reached via the catalog card, existing pattern
  • llms.txt gives Nimble the placeholder "Documentation page", but that hits 21 entries including Tavily (generator fallback when summarization fails). Systemic, worth its own fix.

🤖 Generated with Claude Code


Note

Low Risk
Documentation-only changes to the Nimble integration page and changelog; no runtime, auth, or API behavior changes.

Overview
Updates the Nimble partner integration doc and a one-line changelog style fix so they match the live MCP server and Arcade gateway behavior.

The capabilities section now reflects seven families (including web search agents), explains that the tool picker shows ~two dozen individual tools—not one row per capability—and documents nimble_ prefixes and names like nimble.nimble_search. A callout warns gateway builders to include async polling tools (nimble_task_results, nimble_agents_run_status) with starters, and the page notes which tools write to the Nimble account.

Cost and token usage replaces the old caching narrative with Nimble-attributed retrieval claims, practical MCP token guidance (narrow tool selection, prefer search/extract), and a pointer to Nimble’s CLI skill for high-volume use.

Changelog entry wording changes Partner MCP Server to Partner MCP server for style consistency.

Reviewed by Cursor Bugbot for commit b1d6e47. Bugbot is set up for automated code reviews on this repo. Configure here.

…laims

Follow-up to #1097.

- Add the web search agent capability (Nimble documents six tools; the
  page listed five) and rename "Templates" to "Extract templates" to
  match Nimble's naming.
- Replace the "Caching and cost" section. It described a cache and
  memory layer that "avoids redundant searches and the token cost they
  add" — none of which appears anywhere in Nimble's docs, and which
  contradicts Nimble's own MCP page warning that MCP "can consume
  tokens quickly." Replaced with accurate token-usage guidance and a
  pointer to the CLI skill Nimble recommends for token efficiency.
- Show the real namespaced tool name (nimble.nimble_search) rather than
  a <server-id>.<tool> placeholder, since Nimble prefixes its tools
  with nimble_ and the doubled prefix is surprising.
- Changelog: "MCP Server" -> "MCP server" per STYLEGUIDE.md.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
@vercel

vercel Bot commented Jul 29, 2026

Copy link
Copy Markdown

The latest updates on your projects. Learn more about Vercel for GitHub.

Project Deployment Actions Updated (UTC)
docs Ready Ready Preview Jul 29, 2026 11:38pm

Request Review

I removed this section on the grounds that Nimble's docs didn't support
it. That was wrong — I only checked docs.nimbleway.com. The claims are
published on nimbleway.com/mcp and in Nimble's MCP blog post:

  "Compound accuracy over time through a Proprietary Index and Memory
   that gets smarter with every query"
  "Cut token costs by retrieving exactly what's needed — no redundant
   searches, no parsing raw pages with an LLM"

Restored, with two changes from the original wording. The claims are now
attributed to Nimble rather than asserted as fact, since they're product
claims we can't verify from outside. And the mechanism detail the
original added beyond the source (that the cache stores "sources, search
history, and responses" and the memory layer "extracts key learnings
after each run") is dropped — Nimble publishes "Proprietary Index and
Memory", not that breakdown.

Kept the practical MCP token caveat alongside it. Both are true and in
tension: the marketing claims retrieval efficiency, while Nimble's docs
warn MCP "can consume tokens quickly" and point high-volume users at
their CLI skill. Readers should see both.

Also verified tool names against Nimble's docs: nimble_search is
correct, so nimble.nimble_search stands.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
@thierrypdamiba thierrypdamiba changed the title fix(integrations): correct Nimble page capabilities and token-usage claims fix(integrations): correct Nimble page tool list and naming Jul 29, 2026
Verified against mcp.nimbleway.com/mcp (Nimble 3.4.3, protocol
2025-06-18) via a tools/list handshake rather than published docs.

The server exposes 27 tools, not 6. Nimble's docs describe six
capability families; each ships multiple tools for sync, async, and
status/listing calls. This matters in Arcade specifically, because the
gateway tool picker lists individual tools — someone reading "6
capabilities" then seeing 27 checkboxes has no way to map between them.
Table now carries a per-capability tool count and the total.

Added a callout on the async pattern. Many tools return a task ID and
require polling a companion tool (nimble_task_results,
nimble_agents_run_status). A gateway built with only the starting tools
selected will kick off jobs the agent cannot read back. Worth calling
out since selecting a subset of tools is the normal gateway workflow.

Resolved the open question from the previous commit: the live server has
no Google Maps tools and no nimble_deep_web_search, so the tool names on
nimbleway.com/mcp are stale. Docs were right, marketing page was not.
nimble_search confirmed present, so nimble.nimble_search stands.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The live server returned 27 tools for our account, but that number does
not belong in the docs as a fact. tools/list is auth-gated: requests
without a valid key get 401, so I could not compare what a different
plan or account sees. Our access is a partner-provided trial, which is
exactly the kind of account likely to differ from a customer's.

Arcade's own add-remote-servers guide documents this as expected
behavior: Arcade pre-loads the tools available to the user who
configures the server, and a tool unavailable to an end-user will not
appear via the gateway. So the number is per-account by design.

Now gives the reader the magnitude that matters (expect roughly two
dozen tools, not seven), says the set depends on plan and account, and
points at the guide's admin-user guidance. Capability families stay,
since those are stable across Nimble's docs and the live server.

Also added the read/write split. Most tools are read-only; the agent
create/update/run/delete and template-generation tools write to your
Nimble account, and the server marks agent deletion destructive. Worth
knowing before you drop the whole toolset into a research gateway.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant