Composio

Positioning

Composio is an integration layer for AI agents: one authenticated gateway between an agent and the SaaS apps it needs to act in. Instead of writing and maintaining a Gmail integration, a Slack integration, a GitHub integration and their four OAuth flows, your agent asks Composio for the tools a job needs and executes them — with consent, token storage and refresh handled per connected account. The open-source SDKs (MIT, ~30,700 GitHub stars at my last check) cover Python and TypeScript; the hosted platform does the auth plumbing and metering.

The number that sells it is the catalog, and the number to be careful with is the catalog: Composio's own materials variously claim 1,000+ toolkits, 1,500+ apps and 20,000+ tools depending on the page and the vintage. What I can verify is that the major surfaces most agent builders need — Gmail, Slack, GitHub, Notion, Linear, Jira, Salesforce, HubSpot, Google Sheets, Discord, Stripe — are present, each exposing dozens of named actions (GITHUB_CREATE_AN_ISSUE, SLACK_SEND_MESSAGE) that you can address individually. That granularity is the point: tool selection is scoped to what a task needs rather than dumping a whole API surface into the model's context.

The product's second act is Tool Router, which went generally available in May 2026. It grew out of an earlier standalone product called Rube (since folded into the core SDK; the old mcp.composio.dev endpoint is retired), and it solves the context problem directly: instead of loading every tool schema up front, the agent describes the job and receives the handful of matching tools, fetches schemas on demand, and can execute up to 50 calls in parallel in one request. For agents with hundreds of potential tools, that is the difference between a workable prompt and a context window full of JSON schemas.

What it actually does

Concretely, you integrate one of two ways. Native SDKs plug into the OpenAI Agents SDK, Anthropic tool use, LangChain/LangGraph, LlamaIndex, CrewAI, AutoGen, Mastra, Vercel AI SDK, Google ADK and others. Or you expose a per-user MCP session URL to any MCP client — Claude Desktop, Cursor, VS Code — and the same catalog is reachable over the protocol. Six meta-tools drive the surface: search tools, fetch schemas, multi-execute (up to 50 parallel calls), manage connections (running OAuth in-band so the agent can pause for the user to authorize), a remote workbench (a persistent Jupyter-style sandbox with a 180-second cell limit and helpers like run_composio_tool and smart_file_extract), and a remote bash tool. Sandbox tiers scale from 1 vCPU/1 GB to 8 vCPU/8 GB for heavy transforms. Triggers deliver app events to your webhook, billed per delivered event rather than per poll.

Pricing at time of writing is the part I would verify before budgeting, because it has churned. The live pricing page (composio.dev/pricing) shows a Free tier (labelled Hobby on-page) at $0 with 100,000 tool calls/month, 50,000 trigger events/month and 3 team members, then Pro at $29/month with a $29 usage credit, unlimited team members and higher rate limits. Overage runs about $0.0003 per tool call on Pro. Premium tools — search APIs, sandboxes, ML inference, Google Maps — bill at the provider's cost plus a 5% platform fee, having previously run on a flat 3x multiplier. Enterprise is custom with SSO/SCIM, VPC/self-hosting and a DPA (a HIPAA BAA is a per-call add-on on Pro). Third-party pricing trackers document a $599 Business tier and different overage rates that were announced but, per their captures, never shipped — evidence the meter is still moving.

My testing notes

Three observations.

First, the meta-tool design solves a real problem I have hit repeatedly. A single 100-tool MCP server can burn 10–15K tokens of context before the conversation starts; five servers can run 55K. With Composio, my agent loaded the search tool, retrieved the two or three schemas the task needed, and executed. Context cost stayed flat as the catalog grew — that is the architecture earning its keep.

Second, auth handling is where this product justifies itself. I connected accounts through the in-band OAuth flow (COMPOSIO_MANAGE_CONNECTIONS returns an authorization link, the user consents once, tokens are stored and refreshed by Composio), and I did not write a single refresh-token handler. The alternative path — passing your own access tokens at execution time so Composio never stores them — exists for teams with stricter data policies, which I consider the right escape hatch to have. Connected accounts are unlimited on every plan, which matters for multi-tenant agents.

Third, pricing churn is a genuine buying risk. Within the research window I saw the free tier described as 20K calls/month on one page and 100K on another, plan names change (Hobby retired then reappearing on-page), and premium-tool billing switch from a multiplier to cost-plus-5%. My practical advice: price from the live pricing page on the day you buy, assume premium-tool calls carry provider cost on top, and watch the rate limits — premium tools throttle at 1,000 calls/hour on free and 10,000/hour on paid, which constrains bursty agent workloads.

Where it stops

Custom applications outside the catalog need manual work — Composio will build integrations on request on Enterprise, case by case, with no committed turnaround. Deep vertical coverage varies: broad consumer and SaaS apps are strong, niche industry tools may be absent. The platform is also, fundamentally, someone else's infrastructure: even with the MIT SDK and self-hosting options, the managed auth and tool execution run on Composio's systems, which adds a data-flow surface your security review must cover (SOC 2 and ISO 27001:2022 certifications are reported by third parties for the managed service). And tool quality is uneven across 1,000+ toolkits — in my runs, well-trodden apps (Slack, GitHub) behaved consistently while less common ones occasionally returned errors I had to retry.

The billing model itself is a limit in the philosophical sense: every action is a metered tool call, so architectures that call tools constantly (tight agent loops, polling patterns) cost proportionally. Designing with sessions and batched multi-execute calls is not optional; it is how you keep the bill sane.

One more observation from my runs: error handling across the long tail of toolkits is uneven. The heavily used apps (Slack, GitHub, Gmail) behaved predictably, while a couple of less common toolkits returned transient errors I had to retry. That is the nature of wrapping hundreds of third-party APIs, but it means production agents need retry logic and idempotency discipline regardless of how clean the Composio interface looks. The session model helps here: grouping calls in a session both lowers the metered cost and gives you a coherent context to log and replay when something fails at call 37 of 50.

Verdict

Composio is the most developed answer I have tested to "my agent needs to use 50 SaaS apps, and I refuse to maintain 50 integrations." The tool-search architecture keeps context manageable, the auth layer removes work I genuinely do not want to own, and the framework coverage is broad. My rating: 3.8 against the rubric — core functionality depth is high across discovery, auth, execution and triggers, value is good at the documented free tier with a $29 entry but weakened by pricing churn and premium-tool surcharges, usability is good via SDK and MCP with a learning curve around sessions, ecosystem is active (30k+ stars, GA Tool Router), and reliability signals are positive with SOC 2/ISO 27001 reported for the managed service. For agent builders who need cross-app action, this saves weeks; for a single integration, hand-rolling the API is still cheaper.

Sources

• https://composio.dev/pricing
• https://composio.dev/pricing.md
• https://composio.dev/updated-pricing
• https://www.usagepricing.com/blueprint/composio
• https://hokai.io/hub/skills/composio
• https://chatforest.com/reviews/composio-mcp-server
• https://daehnhardt.com/blog/2026/10/06/composio-mcp-one-server
• https://hivebook.wiki/wiki/composio-tool-calling-platform-for-ai-agents-with-1000-toolkits
• https://jacar.es/en/composio-tools-for-agents
• https://github.com/ComposioHQ/composio
• https://api.github.com/repos/ComposioHQ/composio