
One MCP endpoint that lets an agent act in 1,500+ apps with managed auth.
Composio suits developers running a self-hosted agent such as Hermes Agent or OpenClaw who want it to act in Gmail, Slack, GitHub or a CRM without writing an integration per app. It replaces a folder of single-purpose MCP servers with one endpoint, backed by a repository past 30,000 GitHub stars. Hosted agents like Dots cannot use it.
Tool search over 1,500+ toolkits so an agent loads a handful of tools per job instead of every schema.
The toolkit catalogue and managed-auth plumbing are expensive to replicate, but MCP standardisation lets any app ship its own server, which erodes the aggregator's necessity over time.
Main trade-off
A hosted vendor in the data path of every action unless you self-host or pay for zero data retention.
Hobby is $0 with 100,000 tool calls, 50,000 trigger events and 3 team members a month, no card required. Pro is $29 a month with a $29 usage credit, unlimited members, spend controls and 10,000 requests a minute; past the credit, tool calls bill at $0.0003 and triggers at $0.003. Enterprise is quoted, with SSO/SCIM, customer-managed keys and a BAA included. Self-hosting the open-source core is free.
npx @composio/mcp@latest setup "https://rube.app/mcp" "rube" --client claude
Registered once as a remote MCP server (https://rube.app/mcp) or through the Python/TypeScript SDK. The agent calls COMPOSIO_SEARCH_TOOLS to find the app tools for a job, COMPOSIO_GET_TOOL_SCHEMAS for their parameters, and COMPOSIO_MULTI_EXECUTE_TOOL to run them; COMPOSIO_MANAGE_CONNECTIONS opens the OAuth link for any app not yet connected.
The permissions this connector asks for, as its own documentation states them.
11 actions an agent can call. Open one for its parameters and an example.
Example call
{
"queries": [
{ "use_case": "send a message to a Slack channel", "known_fields": "channel_name:general" }
],
"search_strategy": "auto",
"session": { "generate_id": true }
}Parameters
| Parameter | Type | Needed | Description |
|---|---|---|---|
| queries | array | Required | One or more structured English queries; each has a use_case (what the agent needs to do) and optional known_fields (comma-separated key:value hints such as channel_name:general). |
| search_strategy | stringdefault auto | Optional | auto uses cached plans; tool_search bypasses them and searches tools directly. |
| session | object | Optional | {generate_id: true} for a new workflow, or {id} to continue one. |
| model | string | Optional | The client LLM's model name, used to tune the plan. |
Returns
| Field | Type | Description |
|---|---|---|
| tools | array | Matching tool slugs with descriptions and the toolkits they belong to. |
| plan | string | A recommended execution plan and common pitfalls. |
| session_id | string | Id to pass to the other meta-tools in this run. |
| connections | array | Which toolkits still need an active connection. |
Example call
{ "tool_slugs": ["NOTION_CREATE_PAGE", "HUBSPOT_UPDATE_DEAL"], "include": ["input_schema", "output_schema"] }Parameters
| Parameter | Type | Needed | Description |
|---|---|---|---|
| tool_slugs | array | Required | Slugs copied from a Search Tools result; never guessed. |
| include | arraydefault ["input_schema"] | Optional | input_schema, output_schema, or both. |
| session_id | string | Optional | Workflow session id, if one was issued. |
Returns
| Field | Type | Description |
|---|---|---|
| schemas | object | JSON Schema per tool slug. |
Example call
{
"tools": [
{ "tool_slug": "SLACK_SEND_MESSAGE", "arguments": { "channel": "#general", "text": "Deploy finished" } },
{ "tool_slug": "GITHUB_CREATE_AN_ISSUE", "arguments": { "owner": "acme", "repo": "api", "title": "Flaky test" } }
],
"sync_response_to_workbench": false
}Parameters
| Parameter | Type | Needed | Description |
|---|---|---|---|
| tools | array | Required | 1 to 50 items of {tool_slug, arguments, account?}; account picks one of several connected accounts by alias or id. |
| sync_response_to_workbench | booleandefault false | Required | Save the full response to the remote workbench when it may be large. |
| session_id | string | Optional | Correlates this batch with the earlier search. |
| thought | string | Optional | One-sentence rationale for the batch. |
| current_step | string | Optional | Short step label such as FETCHING_EMAILS. |
| current_step_metric | string | Optional | Progress as done/total, e.g. 3/10 pages. |
Returns
| Field | Type | Description |
|---|---|---|
| results | array | One structured output per tool, in order, each with data or an error. |
Example call
{ "toolkits": [ { "name": "gmail", "action": "add", "alias": "work" } ] }Parameters
| Parameter | Type | Needed | Description |
|---|---|---|---|
| toolkits | array | Required | Items of {name, action, alias?, account_id?}; name is the toolkit slug from Search Tools. |
| toolkits[].action | stringdefault add | Optional | add, list, rename or remove. |
| session_id | string | Optional | Optional; ties the connection to the current run. |
Returns
| Field | Type | Description |
|---|---|---|
| redirect_url | string | The authorisation link to show the user (add). |
| accounts | array | Account ids, aliases and statuses (list). |
Example call
{ "toolkits": ["gmail", "slack"], "mode": "any" }Parameters
| Parameter | Type | Needed | Description |
|---|---|---|---|
| toolkits | array | Required | Toolkit slugs to wait for. |
| mode | stringdefault any | Optional | any or all. |
| session_id | string | Optional | Optional run identifier. |
Returns
| Field | Type | Description |
|---|---|---|
| statuses | object | Terminal state per toolkit. |
Example call
result, error = run_composio_tool('GMAIL_FETCH_EMAILS', {'max_results': 50})
print(len(result['data']['messages']) if not error else error)Parameters
| Parameter | Type | Needed | Description |
|---|---|---|---|
| code_to_execute | string | Required | Python for one cell; state persists across calls; 180-second limit per cell. |
| session_id | string | Optional | Keeps sandbox files scoped to this run. |
| thought | string | Optional | Brief objective for the step. |
Returns
| Field | Type | Description |
|---|---|---|
| output | string | Compact cell output; long results become files under /mnt/files/.composio/output. |
Example call
jq '.data.messages | length' /mnt/files/.composio/output/latest.json
Parameters
| Parameter | Type | Needed | Description |
|---|---|---|---|
| command | string | Required | The bash command; 180-second limit. |
| session_id | string | Optional | Same run as the workbench, so its files are visible. |
Returns
| Field | Type | Description |
|---|---|---|
| stdout | string | Command output. |
Example call
{ "query": "weekly sales report", "limit": 5 }Parameters
| Parameter | Type | Needed | Description |
|---|---|---|---|
| query | string | Required | What the task is. |
| limit | integerdefault 5 | Optional | 1 to 10 results. |
Returns
| Field | Type | Description |
|---|---|---|
| skills | array | Matching Skills with id, slug and title. |
Example call
{ "slug": "weekly-sales-report" }Parameters
| Parameter | Type | Needed | Description |
|---|---|---|---|
| slug | string | Required | Lower-case slug, letters, digits and hyphens. |
Returns
| Field | Type | Description |
|---|---|---|
| skill_markdown | string | The SKILL.md content. |
Example call
{ "action": "create", "title": "Weekly sales report", "skill_markdown": "---\nname: weekly-sales-report\n---\n1. Pull closed deals..." }Parameters
| Parameter | Type | Needed | Description |
|---|---|---|---|
| action | string | Required | create, update or delete. |
| id | string | Optional | Existing Skill id (sk_...) for update or delete. |
| title | string | Optional | Skill title, up to 256 characters. |
| skill_markdown | string | Optional | The complete SKILL.md with YAML frontmatter. |
Returns
| Field | Type | Description |
|---|---|---|
| id | string | The Skill id. |
Example call
{ "tool_slug": "GITHUB_LIST_REPOS", "category": "empty_results", "explanation": "Returned no repositories for an org that has twelve.", "reported_by": "agent" }Parameters
| Parameter | Type | Needed | Description |
|---|---|---|---|
| tool_slug | string | Required | The tool whose latest execution is reported. |
| category | string | Required | empty_results, incorrect_results, transient_failure or missing_capability. |
| explanation | string | Required | 10 to 300 characters, no arguments or payloads. |
| reported_by | string | Required | agent or human. |
Composio is a tool-calling gateway for AI agents: one MCP server (Rube) or SDK exposes more than 1,500 app toolkits, with OAuth consent, token storage and refresh handled per connected account. Agents search for the tools a job needs at run time instead of loading every schema, which keeps context small across hundreds of apps.
Tool search instead of tool loading
The agent describes a job and gets the four to six matching tools back, so a catalogue of hundreds of toolkits never has to sit in its context window.
Managed OAuth per connected account
Consent, token storage, scopes and refresh are handled for every account, and a connected account is free and unlimited on every plan.
Parallel execution
Up to 50 independent tool calls run in one request, each returning structured output.
Remote workbench and bash sandbox
A persistent Jupyter sandbox with a 180-second cell limit runs Python or shell against large tool responses without pulling them into the conversation.
Triggers
Events from connected apps are delivered to a webhook and billed per delivered event, not per poll.
Bring your own tools and credentials
Custom tools and other MCP servers sit beside the catalogue, and self-managed tokens can be passed at execution time so Composio never stores them.
Organisation Skills
Reusable SKILL.md workflows can be searched, loaded and managed by the agent, so a team's recurring jobs are written once.
Give a self-hosted agent hands in your SaaS stack
A Hermes or OpenClaw agent registers the Rube endpoint once and can then file a GitHub issue, post to Slack and update a HubSpot deal from one chat message, with each account authorised through a consent link.
Replace a dozen single-app MCP servers with one
Instead of loading every connector's schema into the agent's context, the agent asks Composio for the few tools a job needs, which keeps prompts small when the toolkit count runs into the hundreds.
Run bulk actions from a sandbox
The remote workbench runs Python against tool results, so an agent can page through thousands of CRM rows or inbox threads and write the summary back without pulling the raw data into the conversation.
React to events in other apps
Triggers deliver new-email, new-issue or new-row events from connected apps to a webhook, so an agent can start a run when something happens rather than on a timer.
| Skill | Works with | Connects by | Checked |
|---|---|---|---|
| Claude · Cursor · Windsurf · VS Code | MCP | 3 Oct | |
| Claude · ChatGPT · Cursor · Windsurf | MCP | 19 Aug | |
| Claude · Codex · Cursor · Windsurf +2 | MCP | 3 Sep | |
| Claude · Cursor | MCP | 26 Oct |
There are three. Hobby is free and stops at its included allowance rather than charging. Pro is $29 a month and includes a $29 usage credit that expires monthly; beyond it, a tool call is $0.0003 and a trigger event $0.003, with add-ons such as a HIPAA BAA at $0.0003 per call and IP allowlisting at $0.0001. Enterprise runs on committed volume with annual invoicing, SSO and SCIM, customer-managed keys and a dedicated SLA.
The Hobby plan includes 100,000 tool calls and 50,000 trigger events a month, three team members, unlimited connected accounts and a 2,000-request-per-minute API limit, with no card required. Meta-tools such as tool search are not counted. The catch is Composio-managed shared OAuth apps: those count only 20,000 calls and 10,000 triggers toward the free allowance, so a serious free deployment registers its own OAuth app per provider.
Zapier MCP exposes Zapier's action catalogue to an MCP client and suits teams already on Zapier. Pipedream Connect and Arcade.dev are the closest gateways with managed auth for developers. If you only need one app, that app's own MCP server, such as the GitHub or Pipedrive servers, is simpler and has no third party in the path. For drawn workflows rather than agent tool calls, n8n self-hosted is the usual choice.
Zapier MCP is the faster start for a non-developer who already runs Zaps, because the actions are the ones Zapier has had for years and the pricing follows the Zapier plan. Composio is built for the agent-first case: tool search that returns a handful of tools per job, parallel execution of up to 50 calls, a sandbox for bulk work, and the option to pass your own tokens or self-host the open-source core. Developers wiring a self-hosted agent pick Composio; a marketing team adding one action to Claude Desktop picks Zapier MCP.
Create an account at app.composio.dev, then either run the one-line MCP setup for your client or paste https://rube.app/mcp into the client's MCP config. The first time the agent needs an app, Manage Connections returns an authorisation link; complete the consent once and the account stays connected. On the SDK path you install the Python or TypeScript package, set the API key and pass the toolkits you want into your framework's tool list.
Composio is a tool-calling layer that sits between an AI agent and the apps it needs to act in. It is a skill in the Hokai sense, not a product a person opens: an agent such as Hermes Agent or Claude Code registers Composio's MCP server once and from then on can search for, authorise and call tools across more than 1,500 app toolkits. The company was founded in 2023 by Soham Ganatra and Karan Vaidya, works out of San Francisco and Bengaluru, and raised a $25 million Series A led by Lightspeed Venture Partners in 2025, taking total funding to twenty-nine million dollars.
The mechanism is what separates it from a folder of single-app MCP servers like the Pipedrive MCP Server. The hosted server, Rube, exposes a small set of meta-tools rather than every app tool at once: the agent describes a job, receives the four to six tools that fit, fetches their schemas, and executes them, dozens of independent calls in one request. Composio runs the OAuth consent and keeps and refreshes the tokens for each connected account, or you pass your own access tokens at execution time and Composio never stores them. A Session is the context a run happens in and is the recommended way to call tools; direct execution outside one is metered separately.
Any MCP client can use it. The documented setups are Claude Desktop, Claude Code, Cursor, VS Code and Windsurf, and the Python and TypeScript SDKs plug into LangChain, CrewAI, the OpenAI Agents SDK and the Vercel AI SDK. Self-hosted personal agents are the natural fit: OpenClaw and Hermes both speak MCP, so one Composio connection gives them hands in Gmail, Slack, GitHub, Notion or HubSpot without a per-app adapter. Hosted bundles such as OpenAI Dots or xAI's Grok Bot bring their own connector layer and cannot take it. Where a workflow is better drawn than coded, n8n and Zapier remain the alternatives, and Composio can sit inside either as the auth layer.
The core is open source under the Elastic License 2.0 and can be self-hosted; the hosted plans start free and meter above the included allowance, with the figures in the pricing section. It runs as a web dashboard, a CLI, an HTTP MCP endpoint and REST and SDK APIs, on any operating system that can reach it.
Development is fast and public. The main repository passed 30,000 GitHub stars in 2026, Rube arrived as the hosted MCP server that solved the context-overload problem of loading hundreds of tool schemas, and the pricing page now lists per-call add-ons for a HIPAA BAA, IP allowlisting and zero data retention, which signals where the company is heading: regulated teams running agents in production. Browse the rest of the skills directory for the single-app connectors it replaces.
Protocol, authentication, permissions and the call contracts on this page are read from Composio's own documentation. The verdict is Hokai's reading of them, not the maintainer's claim.