Composio

Composio is an AI tool-calling and integration platform from Sampark Inc that connects AI agents to more than 1,500 external apps through its MCP Gateway, SDKs and CLI. It solves the authentication, rate-limiting and execution-reliability problems agents hit when taking actions in systems like GitHub, Salesforce, Slack, Gmail, Linear and Notion. Instead of building and maintaining individual integrations, developers get managed connections that handle OAuth flows, credential refresh and error handling, and provider credentials stay out of the agent context. The same setup works from Claude, Cursor, OpenAI Codex, Hermes Agent and other MCP-compatible agents. Its GitHub repository carries 31k stars under an MIT license.

Coverage spans 1,500-plus toolkits and more than 50,000 individual tools across CRM, project management, communication and developer tooling, with managed OAuth, triggers that deliver app events to a webhook, a remote sandbox for running code, and browser tools for websites with no API. Connected accounts are unlimited and free on every tier when you use your own OAuth app or API keys.

Pricing is in USD across three tiers. Hobby is free and hard-capped at 100,000 tool calls and 50,000 trigger events a month with three team members and no credit card required. Pro is 29 dollars a month, adds unlimited team members and spend controls, and includes 29 dollars of monthly usage credit that resets each month and does not roll over. Enterprise is quoted and adds committed volume, annual invoicing, SSO and SCIM, customer-managed keys, an included MSA, DPA and BAA, and a dedicated SLA. Past the included amounts, Pro bills 0.0003 dollars per tool call and 0.003 dollars per trigger event. Customers who signed up before August 15, 2026 keep their earlier pricing through December 31, 2026.

What the 29 dollar tier does not include is the compliance and governance set, which is priced per call on top of the base rate. A HIPAA Business Associate Agreement is an add-on at 0.0003 dollars per tool call, and zero data retention (on select features) and IP allowlisting are 0.0001 dollars per tool call each. All three are included on Enterprise and unavailable on Hobby. Smaller usage add-ons apply when a call runs in the sandbox, outside a session or through proxy execute, and white-labeling the connection page costs 0.30 dollars per connection, once. Connecting through a Composio-managed OAuth application rather than your own includes only 20,000 free tool calls a month and adds 0.0002 dollars per tool call, and on paid plans each managed connection past the first 1,000 costs 0.10 dollars, once.

Composio is best for AI engineering teams that are comfortable working through an SDK, a CLI or an MCP endpoint and want managed tool execution rather than a workflow product. Teams that want to build automations visually without writing code should consider Zapier or Make.com. Teams that need to run the automation on their own infrastructure should consider n8n, although Composio Enterprise can also be deployed in a private VPC or self-hosted. Teams that want a dedicated browser agent for web interfaces with no API should consider Browser Use. Teams that want paid third-party data such as SEO, enrichment and social endpoints from a dedicated pay-per-call data marketplace should consider treg.to or looot.

Current state Q4 2026: Composio Instant launched on October 8, 2026. It gives agents a wallet that pays for more than 100 supported paid tools, such as lead data, web search and email verification, without a separate provider account, at the rate each provider charges with no markup, plus the regular tool-call charge. Subprocessors used by Instant are not covered by the subprocessor list in the trust centre. A new Agent Firewall detects prompt injections in tool responses, and approvals for selected tools are available as an experimental feature. SOC 2 Type II and ISO 27001:2022 are both published as independently audited and certified. The terms updated October 17, 2026 state that tool-call payloads are never used to train AI models, while de-identified execution metadata, such as which tool ran and its status, may be used to train machine learning models.