How Tiny Fish's MCP server adds free web search, page fetch, and paid browser automation to DeepSeek Harness or any coding agent.
What is Tiny Fish and why does a coding agent need it? #
Tiny Fish is web infrastructure built for AI agents. It sits between your agent and the live web, handling the messy parts of browsing (rendering pages, clicking through forms, managing logins, dodging bot detection) and handing your agent back clean, structured results. Most agentic coding harnesses, including DeepSeek Harness, ship with basic web search and a fetch tool that scrapes public pages anonymously. Neither can click a button, fill out a form, or log into an account. Tiny Fish fills that gap by giving an agent an actual browser it can drive, plus a model that can plan multi-step web tasks on its own.
What tools does Tiny Fish actually provide? #
Tiny Fish exposes four capabilities through a single API, reachable from any MCP-compatible agent:
Web search returns live search results, structured as titles, links, and snippets, rather than a raw page dump.
Fetch takes a URL, renders it in a real browser (not a bare HTTP request), and returns the page as clean text, stripped of menus, ads, and clutter.
Web agent takes a plain-English goal, such as “log into this account and read the dashboard,” and handles the navigation, clicking, and form-filling itself, returning structured JSON. It runs on Tiny Fish’s own model, called Marco, and pricing is per step with the model cost included.
Other agents start typing. Remy starts asking. #
Scoping, trade-offs, edge cases — the real work. Before a line of code.
Browser gives you a cloud browser instance you can drive yourself with Playwright. It includes anti-bot protections built into the browser engine and a credential vault, so an agent can authenticate into a site without your actual password ever being exposed to the model.
Search and fetch are free on every plan, at any volume. The web agent and browser products are the paid tier, billed per step, and that’s what the Tiny Fish wallet (funded by sign-up credits) is for.
Why can’t DeepSeek Harness do this out of the box? #
DeepSeek Harness, the open-source agentic coding harness built around DeepSeek’s models, already ships with two web tools: a search tool that runs DeepSeek’s own search inside a model turn, and a fetch tool that downloads a public page and hands the model its text. Both are read-only. The fetch tool is deliberately anonymous: it sends no cookies, no credentials, and doesn’t execute JavaScript or render the page the way a real browser does. That’s fine for pulling information off a public page, but it’s useless for anything behind a login, a dashboard, or a multi-step UI flow. DeepSeek Harness does have experimental browser-use and computer-use plugins in development, but they aren’t installed by default and require wiring up your own local browser. Tiny Fish is a way to get that capability now, without building the browser automation layer yourself.
How do you add Tiny Fish to DeepSeek Harness? #
DeepSeek Harness is built around a plugin architecture. The desktop app is essentially a host, and nearly everything, including model routing, web search, and the terminal, runs as a plugin package. Each profile has a patch file where plugins get turned on, off, or configured, alongside a plugin manager in the UI. That architecture matters here because adding Tiny Fish isn’t a code change, it’s a configuration change.
Tiny Fish runs as an MCP server, and DeepSeek Harness includes an MCP client plugin. To connect the two:
- Open the desktop profile’s patch file.
- Add the MCP client plugin.
- Give it a server name and point it at the Tiny Fish endpoint.
- Pass your Tiny Fish API key as a header, ideally read from an environment variable rather than hardcoded, so you can share the config file without leaking credentials.
- Restart the app and start a new session.
Once restarted, the four Tiny Fish tools (search, fetch, web agent, browser) show up alongside the harness’s existing tools. The same MCP server works with other agentic coding tools, including Claude Code, Codex, and Cursor, using similarly simple installation steps.
What can the agent actually do once it’s connected? #
With Tiny Fish wired in, an agent stops being limited to reading public pages and starts being able to act on the web. Demonstrated use cases include:
- Logging into a site using test credentials, reading the authenticated page content, and logging out again, returning the result as JSON, with a live view of the browser session as it clicks through each step.
- Navigating to a filtering and sorting interface (in one case, Hugging Face’s model listings), applying filters, sorting results, and extracting a list that matches specific numeric criteria pulled from the page, down to catching naming inconsistencies between a model’s name and its actual listed parameter count.
- Setting up a monitor that checks a specific page (a software release page, in the demonstrated case) on a recurring schedule and only alerts when something meaningfully changes, rather than reporting back every time it checks.
- Saving a login session once inside a Tiny Fish browser session, so the agent can reuse that authenticated session later without ever being handed the actual password.
Remy is new. The platform isn't. #
Remy is the latest expression of years of platform work. Not a hastily wrapped LLM.
These are tasks that neither the DeepSeek Harness’s built-in search nor its fetch tool can perform, because they require interacting with page elements, not just reading rendered text.
Is Tiny Fish worth the cost for coding agents? #
For most read-only tasks, search and fetch cost nothing, so there’s little downside to wiring them in. The cost question only matters once you move to the web agent or browser products, which bill per step. In practice, simple tasks like a login-and-read flow or a filtered, sorted data pull came in at well under a dollar in agent steps. Longer-running tasks, like multi-step scraping or monitors that poll a page every few hours, take a few minutes to resolve and are better run in the background rather than watched synchronously, since an agent that keeps re-checking a pending job can burn tokens just waiting. Tiny Fish also runs a credit-matching offer for users switching from Firecrawl: submitting a recent Firecrawl invoice gets matched in Tiny Fish credits, which is worth noting if you’re already paying for a comparable scraping service and want to test the browser-automation side without spending more upfront.
Frequently Asked Questions #
Does Tiny Fish replace DeepSeek Harness’s built-in web search?
Not exactly. DeepSeek Harness’s native search and fetch tools still work and cost nothing extra beyond normal model usage. Tiny Fish’s search and fetch are also free, but render pages in a real browser rather than anonymous HTTP requests, which can produce cleaner results, and they enable the paid web agent and browser tools that the built-in tools don’t have at all.
Can Tiny Fish log into websites without exposing my password to the AI model?
Yes. You can log into a site once yourself inside a Tiny Fish browser session and save it as a profile, or use the credential vault. The agent reuses that authenticated session on later runs without the password ever being passed to the model.
Which agentic coding tools support Tiny Fish besides DeepSeek Harness?
Tiny Fish runs as a standard MCP server, so it connects to any MCP-compatible agent, including Claude Code, Codex, and Cursor, using the same kind of configuration used for DeepSeek Harness.
What’s the actual cost of using the Tiny Fish web agent?
Search and fetch are free at any volume. The web agent and browser products bill per step, with Tiny Fish’s own model (Marco) cost included in that per-step price. Demonstrated tasks like a login-and-logout flow or a filtered data pull from a model listing site each cost well under a dollar.
Do I need to self-host a browser to use Tiny Fish’s browser automation?
No. The browser product gives you a cloud browser instance with built-in anti-bot protection that you can drive with Playwright, so there’s no need to run or maintain your own local browser infrastructure.