Giving AI Coding Agents Browser Access: A Guide to Tiny Fish Setup, Pricing, and DeepSeek Harness Integration

Most AI coding agents can search the web and read public pages, but they cannot click buttons, fill out forms, or log into accounts. This gap limits what agents can do with live web data. Tiny Fish, an MCP-based web infrastructure layer, closes that gap by giving agents a real browser they can drive, along with a model that plans multi-step web tasks. This article explains what Tiny Fish provides, why DeepSeek Harness needs it, how to set it up, and what it costs.

Tiny Fish is web infrastructure built specifically for AI agents. It sits between an agent and the live web, handling the difficult parts of browsing: rendering pages, clicking through forms, managing logins, and avoiding bot detection. It returns clean, structured results the agent can use directly. Its capabilities are reachable from any MCP-compatible agent through a single API.

Tiny Fish exposes four tools. Web search returns live results as titles, links, and snippets rather than raw page dumps. Fetch takes a URL, renders it in a real browser instead of making a bare HTTP request, and returns clean text stripped of menus, ads, and clutter. Web agent accepts a plain-English goal, such as logging into an account and reading a dashboard, then performs the navigation, clicks, and form filling itself, returning structured JSON. It runs on Tiny Fish's own model, Marco. Browser provides a cloud browser instance you control with Playwright, including built-in anti-bot protections and a credential vault that lets an agent authenticate without exposing your password to the model.

DeepSeek Harness, the open-source agentic coding harness built around DeepSeek's models, includes two web tools: a search tool that uses DeepSeek's search within a model turn, and a fetch tool that downloads a public page's text. Both are read-only. The fetch tool sends no cookies or credentials and does not execute JavaScript or render pages like a real browser. This works for reading public information but cannot handle logins, dashboards, or multi-step interfaces. Experimental browser-use and computer-use plugins exist but are not installed by default and require a self-managed local browser. Tiny Fish offers this capability without building the automation layer yourself.

DeepSeek Harness uses a plugin architecture, so connecting Tiny Fish is a configuration change rather than a code change. Open the desktop profile's patch file and 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 so the config file can be shared safely. Restart the app and start a new session. The four Tiny Fish tools will then appear alongside the harness's existing tools.

With Tiny Fish connected, an agent can act on the web rather than only read it. Demonstrated use cases include logging into a site with test credentials, reading authenticated content, and logging out, returning results as JSON with a live view of each step. An agent can navigate a filtered listing interface, such as model listings on Hugging Face, apply filters and sorting, and extract entries matching numeric criteria, even catching inconsistencies between a model's name and its listed parameter count. An agent can also set up a recurring monitor on a page, such as a software release page, and alert only when something meaningful changes. Finally, a login session saved once inside a Tiny Fish browser can be reused later without the agent ever receiving the password.

Search and fetch are free at any volume on every plan, so there is little downside to enabling them. The web agent and browser products are the paid tier, billed per step with model cost included. Sign-up credits fund the Tiny Fish wallet for these paid tools. In practice, simple tasks such as a login-and-read flow or a filtered data pull cost well under a dollar in agent steps. Longer tasks, such as multi-step scraping or monitors polling a page every few hours, take minutes to resolve and are better run in the background, since an agent repeatedly re-checking a pending job can consume tokens while waiting. Tiny Fish also offers a credit-matching promotion for users switching from Firecrawl: submitting a recent Firecrawl invoice is matched in Tiny Fish credits.

Because Tiny Fish runs as a standard MCP server, it is not limited to DeepSeek Harness. The same server connects to other MCP-compatible agentic coding tools, including Claude Code, Codex, and Cursor, using similarly simple installation steps. This makes it a portable web-access layer across different development environments.

Does Tiny Fish replace DeepSeek Harness's built-in search? Not exactly. The native tools still work at no extra cost beyond normal model usage. Tiny Fish's search and fetch are also free, render pages in a real browser, and enable the paid web agent and browser tools that the built-in tools lack. Can Tiny Fish log in without exposing my password to the model? Yes. You can log in once inside a Tiny Fish browser session and save it, or use the credential vault, so the agent reuses the authenticated session without receiving the password. Which tools besides DeepSeek Harness support it? Any MCP-compatible agent, including Claude Code, Codex, and Cursor. What does the web agent cost? Search and fetch are free. Web agent and browser products bill per step, with model cost included; demonstrated tasks cost well under a dollar. Do I need to host my own browser? No. The browser product provides a cloud instance with built-in anti-bot protection that you drive with Playwright.

DeepSeek Harness and similar coding agents are capable at reasoning and code generation, but their web access has been limited to reading public pages. Tiny Fish extends that access to interactive, authenticated, and multi-step web tasks through a simple MCP configuration. Read-only search and fetch can be used at no cost, while paid per-step automation covers tasks that require real interaction. For developers who need agents to work with live, login-protected, or dynamic web content, Tiny Fish offers a practical path without building browser automation infrastructure from scratch.