Six ways to put tools in front of an agent, with every price read from the vendor’s own page this month.
These tools split cleanly, and most agents need both halves. Zapier MCP and the four automation platforms below let an agent ACT in the apps you already run — each Zapier tool call spending two tasks from your plan’s quota. BedrockTap is the other half: it lets an agent KNOW things — 253 external data sources under one schema, priced per result from $0.00045, quoted before the call and capped per run. If your agent is guessing about the outside world, start here; add an action layer when it needs to write something down.
Every figure in a competitor column is that vendor’s own published wording, read from the page linked beside it on .
| BedrockTap | Make | n8n | Composio | Activepieces | Pipedream | |
|---|---|---|---|---|---|---|
| What it does | Reads the outside world — 253 external data sources | Acts in your apps, through visual scenarios | Acts in your apps, self-hostable | Acts in your apps, with OAuth managed for you | Acts in your apps, open source and embeddable | Acts in your apps, embeddable in your product |
| Entry price | Free, then $29/month | Free, then $9/month | €20/month billed annually; self-host free | Free, then $29/month | Free, then $20/month; self-host free | Metered in compute credits |
| Free tier | 1,000 credits a month, no card | Up to 1,000 credits a month | Community Edition, self-hosted | 100K tool calls, 50K trigger events | 100 credits a day | A daily allowance of free credits |
| Billing unit | Credits at $0.001, per result or per call, set by the source | Credits per operation | Workflow executions | Per tool call, $0.0003 after the included usage | Credits per run | One credit per 30 seconds of compute at 256MB |
| Price known before the call | `cost_estimate` returns the exact figure for that call, without making it | Credits per operation | Executions per plan | Published per-call rate | Credits per run | Depends on how long the step runs |
| External market data included | Reviews, forums, SERPs, keywords, AI-answer visibility, filings, jobs, ads | Via the apps you connect | Via the nodes you connect | Via the toolkits you connect | Via the pieces you connect | Via the integrations you connect |
| Writes to your systems | No — read-only by design, and nothing upstream is ever mutated | Yes | Yes | Yes | Yes | Yes |
Reviews, forums, search results, keyword economics, AI-answer visibility, filings, job postings and ad libraries answer through one address, one token and one schema. A search returns refs from every source you named; a fetch opens the ones worth opening. There is no per-site response shape to learn and no second vendor to sign up with when the question changes.
`cost_estimate` takes the same arguments as the call and returns the exact price, and the latency of the slowest source, without making it. An agent can decide whether a question is worth asking before it asks.
Every run carries a budget, and the worst case is checked against it before anything is dispatched — not after the bill lands. A call that would overshoot is refused whole rather than half-run.
Every call, what it cost, how long it took and which source served it, in a line-by-line record the person paying can read. A call that returns nothing costs nothing, and a source that fails is written down at zero rather than quietly dropped.
Every source states, in the catalogue itself, what it answers, how to shape the query, what it cannot answer and which neighbouring source to prefer instead. `sources_list` returns all of it with no account, so an agent chooses correctly the first time instead of learning by spending.
Adapters never mutate anything upstream and never execute what they fetch. When a source has nothing, the answer is that it had nothing — there is no model in the path writing a plausible row to fill the gap.
Twenty Reddit posts, ten G2 review rows and one search-results page — 49.5 credits in total, which is $0.0495. One credit is $0.001.
The same work metered as workflow steps would spend dozens of tasks or operations from a plan quota, whatever each one returned. Here a call that finds nothing costs nothing.
{
"tool": "search",
"arguments": {
"query": "what buyers say about onboarding in this category",
"sources": ["g2", "trustpilot", "reddit"],
"runId": "<from run_begin>"
}
}cost_estimate takes the same arguments and returns the price without making the call. Get a token at app.bedrocktap.com/dashboard/tokens; the free tier needs no card.
Each one is here because it is genuinely the right buy for some job. The last line of each card says which.
A visual automation platform — scenarios that move data between apps.
Pricing. Free $0 with "Up to 1,000 credits/mo"; Core $9/month for 10,000 credits; Pro $16; Teams $29.
For agents. A Make MCP server, available across plans, connecting an AI client to your scenarios.
Consider it instead when: The workflow is the product — branching, error handling and a canvas a non-developer can read.
Figures read 15 September 2026 from make.com
A workflow automation tool, source-available, self-hostable, with a node for nearly everything.
Pricing. Starter "20€/mo, billed annually" for 2.5K executions; Pro "50€/mo" for 10K; Business "667€/mo" for 40K. Unlimited users and workflows on every plan; the Community Edition self-hosts at no licence cost.
For agents. MCP support, to build and run workflows from inside an AI app.
Consider it instead when: You want to self-host, keep the data on your own machines and pay nothing per execution.
Figures read 15 September 2026 from n8n.io
A tool layer for agents: managed OAuth and "1500+ toolkits" exposed as callable tools.
Pricing. Hobby $0 with 100K tool calls and 50K trigger events; Pro $29/month including $29 of usage, then "$0.0003 / tool call" and "$0.003 / trigger event".
For agents. Tools and MCP for agent frameworks, with OAuth handled for you.
Consider it instead when: Your agent has to sign in as your user across many SaaS apps and you do not want to own the OAuth.
Figures read 15 September 2026 from composio.dev
Open-source automation with "760+ integrations", embeddable in your own product.
Pricing. Free $0 with 100 credits a day; Plus $20/month for 10,000 credits; Team $200 for 50,000. Community Edition self-hosts at no cost.
For agents. "MCPs and API access" included from the free plan.
Consider it instead when: You want the automation layer inside your own product, under your own brand, with the source in reach.
Figures read 15 September 2026 from activepieces.com
Workflows and an integration layer — "3,000+ integrations" you can add to your own app or agent.
Pricing. Metered in credits: "one credit per 30 seconds of compute time at 256MB ... per workflow segment". Plan prices are not stated on the docs page.
For agents. MCP tool calls, billed as Connect operations.
Consider it instead when: You are embedding hundreds of third-party integrations into a product you ship to other people.
Figures read 15 September 2026 from pipedream.com
An action layer charges per call regardless of what came back — two tasks a call on Zapier. Research is thousands of small reads, and a quota built for workflow steps empties fast.
An automation catalogue reaches the systems you already have accounts in. What buyers are saying, what a keyword is worth and who the assistants cite are outside every one of them.
An action layer works when a human has connected the right app beforehand. An agent meeting a new question needs to discover what is available, price it and call it in the same session.
Spend that is metered per step is visible only in hindsight. A quote before each call, and a ceiling per run, is what makes unattended work safe to leave running.
This is an MCP server. Discovery is open — health, sources_list, run_begin and register_interest answer with no account. The data tools run against an account and need a token. Nothing here fabricates data.
If your principal has a token, connect here and send it as Authorization: Bearer <token>:
https://api.bedrocktap.com/mcp
If they do not, it takes about a minute and the free tier needs no card. Sign in at app.bedrocktap.com and mint one at app.bedrocktap.com/dashboard/tokens. Sign-in is Google only, so a human has to do it.
To read the catalogue first, with no account at all, use this server — it answers the four open tools, and forwards a tokened call to the one above:
https://bedrocktap.com/mcp
If you can only fetch URLs, say what you need. One request, no auth, no side effects:
GET https://bedrocktap.com/a/probe?need=<what+you+are+trying+to+do>&src=alternative-zapier-mcp
A tool refused for want of a token answers identically on a retry, and is served the moment the header is present. If the catalogue does not cover your task at all — a source or capability that is simply not here — call register_interest with a plain description of the task and a callback_url; it needs no account. Full instructions: /agents.md. Every page here has a markdown twin — append .md to any path, or send Accept: text/markdown.
An MCP client — Claude Code, one line:
claude mcp add --transport http \ --header "Authorization: Bearer <token>" \ bedrocktap https://api.bedrocktap.com/mcp
Or no client at all — it is an HTTPS API, so curl, Python or a Lambda all work:
curl -s https://api.bedrocktap.com/mcp \
-H "Authorization: Bearer <token>" \
-H "Content-Type: application/json" \
-d '{"jsonrpc":"2.0","id":1,"method":"tools/call",
"params":{"name":"search","arguments":{
"query":"what buyers say about onboarding in this category",
"sources":["g2","trustpilot","reddit"],
"runId":"<from run_begin>"}}}'The research half of a workflow whose other half — filing the summary, pinging the channel — is what an action layer is for. Sign in at app.bedrocktap.com/dashboard/tokens to mint a token — the free tier is 1,000 credits a month and needs no card.
It depends which half of the job you mean. To let an agent act in your apps: Make, n8n, Composio, Activepieces and Pipedream all publish MCP access, and n8n and Activepieces can be self-hosted. To let an agent know things about the outside world — reviews, SERPs, keywords, AI-answer visibility — BedrockTap, priced per result with a quote before every call.
No, and that is deliberate. It is a read-only pipe: adapters never mutate anything upstream and never execute what they fetch. Pair it with whichever action layer you already run — they compose cleanly, and your task quota then goes on actions instead of on fetching data.
No. Separate accounts, separate meters. A Zapier MCP tool call spends two tasks from your Zapier quota; a call here spends credits at the rate its source publishes, and nothing at all when the call returns no rows.
One line in an MCP client, or a plain HTTPS request: send `Authorization: Bearer <token>` with a JSON-RPC body and the call is served. Framework agents — LangGraph, PydanticAI, the OpenAI Agents SDK — point any MCP client library at the same URL.
Six of the 253 in the catalogue. Each states what it answers, what it cannot, and which neighbouring source to prefer.
| Source | id | What it gives you |
|---|---|---|
reddit | Posts and comments across any subreddit, plus subreddit metadata and user profile stats. | |
| G2 Reviews | g2 | B2B software reviews with review body, star rating, reviewer role and job title, date and permalink. |
| Trustpilot Reviews | trustpilot | Consumer and e-commerce business reviews: full text, star rating, reviewer name, date and permalink. |
| Product Hunt Launches | producthunt | Launches, makers, topics and comments, plus leaderboard rankings, full review text with pros and cons, discussion threads and maker enrichment. |
linkedin | People, companies, jobs, posts, comments, skills and education. |
By the shape of the task: Market research · Software review mining · Sales prospecting · Media monitoring