Up And Running With Composio

Composio is a hosted catalog of authenticated tool integrations for LLM agents — you connect an account once (OAuth, API key) and your agent gets typed actions against that service without you writing the client. We added it to augment-it to evaluate whether its LinkedIn toolkit could speed up our common enrichment tasks. Short answer on LinkedIn below; it is not what we hoped, and the why is worth recording.

Installation

In the augment-it pnpm workspace root:
zsh 
pnpm add -w @composio/core @composio/claude-agent-sdk @anthropic-ai/claude-agent-sdk
The -w is required — pnpm refuses a bare pnpm add at a workspace root (ERR_PNPM_ADDING_TO_ROOT) to stop you accidentally hanging a dependency off the root instead of a package. These three are root-level, so -w is correct.
⚠️Two unmet peer-dependency warnings (install still succeeds)
@anthropic-ai/claude-agent-sdk
wants
zod@^4.0.0
(workspace has
3.25.76
) and
@anthropic-ai/sdk@>=0.93.0
(resolved
0.69.0
). pnpm does
not
auto-upgrade peers. These are warnings, not failures — bump
zod
and
@anthropic-ai/sdk
only if the agent SDK throws at import or type-check time (zod v3→v4 had breaking changes).
It installs into the workspace node_modules, not globally.

The finding: Composio's LinkedIn toolkit is a write surface, not a read surface

This is the part to remember. Our augment-it context-v corpus wants LinkedIn for reads on arbitrary entities — find a stranger's profile URL, location, headline, and current role given only a name. That need recurs across nearly every enrichment design:
  • Profile discovery for any entity — Common-Six-Social-Packs, Entity-Profile-Augmentation-Workflow, Connector-Inventory-and-Per-Record-Palette
  • Location inference over a network — LinkedIn-Network-Explorer-For-Curated-Invites, Joined-People-UI-and-the-Network-First-Pivot
  • Personal-network extraction — list your own connections, then enrich
  • Sparse-person resolution — Sparse-Person-Enrichment-Surface
Composio's LinkedIn toolkit is the official LinkedIn OAuth2 API. Its ~22 actions are overwhelmingly write/marketing: create post, create comment, create article, upload image/video, ad-targeting facets, audience counts, organization-page statistics. The only reads are "Get my info" (your own profile) and "Get person profile" by member ID — and per LinkedIn's API policy that second one only returns the authenticated member or members who have explicitly authorized your app.
⚠️There is no people-search, and no way to fetch a stranger's location/headline
This is not a Composio limitation — it is LinkedIn policy. Our own
LinkedIn-Network-Explorer-For-Curated-Invites
already documented the wall: the
/me/connections
endpoint was deprecated ~2015, there is no public search API, and scraping gets accounts banned.
Composio does not change that reality.
Enrichment, location inference, profile discovery, and sparse-person resolution correctly stay on the search-inference path (SearXNG / Tavily / SerpApi).
One architectural consequence: Connector-Inventory-and-Per-Record-Palette slots a future linkedin-public connector to serve the search.social.linkedin intent. Composio cannot fill that slot — that intent is a people-search, which the official API does not offer.

Where Composio's LinkedIn toolkit does earn its place

The outbound + client-reporting side — barely touched in our corpus today, but real consulting work:
✅Legitimate, API-blessed, no scraping risk
  • Client org-page analytics for reporting — Get organization page statistics, Get share statistics, Get network size, Get company info give authoritative impressions / clicks / follower-growth for a client org the operator manages. Good raw material for a VC-client dashboard.
  • The actual outreach step of the invite workflow — LinkedIn-Network-Explorer ends in inviting a curated list. Composio can post the announcement to the operator's or client's company page and manage comments — turning "build the list" into "build the list and publish."
  • Own identity / managed orgs — Get my info + Get company info (orgs where you hold posting rights).
These scope cleanly to our per-client client_access[] / workspace model, because they act as the operator's own authenticated identity — not arbitrary-entity data.

Recommendation

  • Do not wire Composio LinkedIn to accelerate enrichment — it can't, and pretending it can would corrupt the connector-inventory model.
  • Do use it narrowly for client org-page analytics and outbound posting.
  • If you want Composio to speed up augment-it's actual common tasks, the high-value toolkits are not LinkedIn — they are the CRM/comms connectors (Gmail, Google Calendar, a CRM alongside the existing Decile integration) that map onto the land-back side of the augmentation pipeline, where official-OAuth connectors genuinely shine.
📝The general lesson
When evaluating any "agent gets superpowers" integration catalog, check the
read/write asymmetry of the underlying API
before the toolkit's marketing. A platform that locks down reads (LinkedIn) exposes a toolkit that is necessarily lopsided toward writes — no wrapper can give you data the source API refuses to return.