Jonathon's AI Wiki

TinyFish — Web Infrastructure APIs for AI Agents

7 min read

Source: May 4 2026 free-tier announcement · MCP integration docs · MarkTechPost April 14 2026 launch coverage

TinyFish is a four-product web-infrastructure platform for AI agents — Search, Fetch, Browser, and Agent — under one API key. It launched the unified platform in April 2026 with a $47M Series A from ICONIQ, then made Search and Fetch free across every surface on May 4 2026. The pitch: own the full Chromium stack end-to-end so the two primitives every agent needs (search the web, read a page) become free utilities like DNS or routing — and the heavyweight surfaces (browser sessions, multi-step Agent) sell on raw infrastructure quality rather than on bundling free things.

Key Takeaways

  • Search + Fetch are free across every surface (REST API, MCP server, Python and TypeScript SDKs, CLI, Skill, plus integrations: OpenClaw, Hermes Agent, n8n, Dify, LangChain, CrewAI). One sign-up at agent.tinyfish.ai, one key. Free-tier rate limits: 5 queries/min for Search, 25 URLs/min for Fetch. Failed Fetch URLs don’t count against quota.
  • Paid surfaces: run_web_automation = 1 credit/step, create_browser_session = 1 credit per 4 browser-minutes, batch_create runs up to 8 simultaneous automations. 500 free credits on signup, no card required.
  • Full Chromium rendering + 28 C++-level anti-bot mechanisms — handles JS-heavy SPAs, paywalls, and CDN-fronted sites that defeat naive requests.get. Stealth Chrome session via CDP.
  • Latency edge: P50 488ms search vs ~2,800ms competitors. Sub-250ms browser cold start vs 5–10s competitors.
  • CLI > MCP for heavy work: vendor benchmarks show 87% token reduction and 2× higher task-completion rate when using TinyFish CLI + Skill instead of the MCP server. CLI writes results to filesystem; agent reads only what it needs. MCP dumps everything into context.
  • Single session identity across an entire workflow — same IP, fingerprint, cookies. Separate-tool stacks appear as multiple unrelated clients to target sites, raising detection risk and failure rates.
  • In-house full stack — Search, Fetch, Browser, and Agent all built by the same team. Compared in launch coverage to Browserbase (uses Exa for search, layer not proprietary) and Firecrawl (agent endpoint reliability issues).
  • Customers include Google, DoorDash, Cigna, Volkswagen, Grubhub, NEC. Reports 40M+ agent operations processed and 99.99% platform uptime.

The four products

Custom-Chromium-backed web search returning structured JSON. Rank-stable across calls. ~488ms P50. Designed as the discovery step in front of Fetch, the retrieval layer in a RAG pipeline, or a verification step inside a multi-turn agent run.

Fetch

Renders any URL in a real browser (full JS, SPAs, dynamic content) and returns clean Markdown / JSON / HTML — strips nav, scripts, cookie banners, ads, footers. Up to 10 URLs per call. Failed URLs free. The stated goal is cutting token cost on every call by stripping noise before it enters the agent’s context.

Browser

Managed stealth Chrome session via Chrome DevTools Protocol. Sub-250ms cold start. 28 anti-bot mechanisms built at the C++ level rather than via JavaScript injection. 1 credit per 4 browser-minutes.

Agent

Autonomous multi-step web workflows — navigates sites, fills forms, clicks through flows, returns structured results without manually scripted steps. run_web_automation streams real-time progress; run_web_automation_async returns a run ID immediately for polling. 1 credit per step. batch_create can launch up to 8 simultaneous automations.

Why it matters for this wiki

Two angles:

Better fallback for WebFetch-blocked sites. The Karpathy-wiki research workflow already documents that WebFetch returns 403 on Substack, claude.com/blog, and most CDN-fronted changelogs — Tavily extract is the canonical fallback. TinyFish Fetch is a stronger fallback still: full Chromium rendering with C++-level anti-bot bypass handles sites that even Tavily extract sometimes misses. And the rate limit (25 URLs/min free) covers solo wiki ingest comfortably.

Calibration on the CLI-vs-MCP trade-off. TinyFish’s own benchmarks (87% token reduction, 2× completion rate when using CLI + Skill over MCP) are concrete numbers for the same trade-off explored in prompt caching for agencies. The pattern: dump-to-filesystem-and-let-the-agent-read keeps context windows clean across multi-step workflows; sequential MCP round-trips blow context. The lesson generalizes — for any heavyweight web-infra MCP, consider whether a CLI + Skill alternative exists for production traffic.

Implementation

Tool/Service: TinyFish (https://agent.tinyfish.ai) — REST + MCP + CLI + SDKs

Setup:

  1. Sign up at agent.tinyfish.ai (no credit card)
  2. Grab the API key from the dashboard
  3. Pick a surface: REST, MCP, CLI (npm install -g @tiny-fish/cli), Python (pip install tinyfish), TS (npm install @tiny-fish/sdk), or the open-source Skill at github.com/tinyfish-io/tinyfish-cookbook

Cost: Search free (5 q/min). Fetch free (25 URLs/min). Browser 1 credit per 4 minutes. Agent 1 credit per step. 500 credits free on signup. Higher rate limits at tinyfish.ai/pricing.

MCP install for Claude Code:

claude mcp add --transport http --scope user tinyfish https://agent.tinyfish.ai/mcp \
  --header "Authorization: Bearer $TINYFISH_API_KEY"

OAuth 2.1 is the documented auth path; Bearer token via Authorization header is also accepted by the endpoint. --scope user keeps the key out of any project-committed .mcp.json.

Manual JSON config (Claude Desktop / Cursor / Windsurf):

{
  "mcpServers": {
    "tinyfish": { "url": "https://agent.tinyfish.ai/mcp" }
  }
}

Restart the client after install. First tool invocation either succeeds via the Bearer header or opens a browser OAuth flow that completes automatically when both claude.ai and agent.tinyfish.ai are signed in.

Integration notes:

  • Works with Claude Code, Cursor, Codex, ChatGPT desktop, OpenClaw, Hermes Agent, Cline, Goose, Antigravity, n8n, Dify, LangChain, CrewAI
  • Vendor recommends CLI + Skill over MCP for heavy multi-step work — 87% token reduction, 2× completion rate on complex tasks
  • MCP path best for: discovery, exploration, single-shot fetches inside an interactive session
  • CLI path best for: production multi-step workflows, batch processing, anything where context cleanliness matters across many turns

Try It

  1. Replace WebFetch fallback for blocked CDN sites. When Substack / claude.com/blog / Medium return 403 to WebFetch, route through mcp__tinyfish__fetch_content instead of Tavily extract. Full Chromium handles JS-rendered + paywalled content.
  2. Drop a Search call into a research routine. Swap one Tavily search for mcp__tinyfish__search and compare result quality and latency on a topic that crosses recent news.
  3. Try the Skill, not just the MCP. For a multi-source compile (e.g., 5 X-bookmarks + 3 YouTube transcripts), install the TinyFish Skill from github.com/tinyfish-io/tinyfish-cookbook/tree/main/skills/use-tinyfish and let the agent route via CLI. Compare token usage to a pure-MCP run on the same workload.
  4. Wire it into Hermes for unattended scraping. Since Hermes natively supports TinyFish, set up a daily watchlist refresh that uses fetch_content to snapshot competitor pricing pages — same routine Clawdbot — Competitive Intelligence (internal)|Clawdbot (internal — references weomarketly-vault topics) runs but with the full-Chromium fetch layer.
  • Essential MCP Servers — TinyFish slots into the high-value MCP server list alongside Tavily, Exa, and Brave
  • AITmpl Curated Templates — already names TinyFish in the AITmpl Anthropic-OSS-program-backed catalog
  • Prompt Caching for Agencies — TinyFish’s CLI-vs-MCP 87% token reduction is a directly comparable data point for context-window economics
  • Clawdbot — Competitive Intelligence (internal)|Clawdbot — Competitive Intelligence (internal) — Clawdbot uses Apify + Brave for similar competitive scraping; TinyFish is the in-house-stack alternative
  • Hermes Agent — Hermes natively integrates TinyFish as a web-infra layer
  • Cowork + Apify Scraping Recipe — operator-track scraping; TinyFish is the developer-track equivalent
  • Agent Workflow Patterns — the sequential-vs-parallel-vs-evaluator framework that TinyFish’s batch_create and Agent endpoints map onto

Open Questions

  • How does the Bearer-header auth path coexist with OAuth 2.1 in Claude Code’s MCP client over time? Documented auth is OAuth; Bearer also works but the lifecycle (token rotation, expiry) isn’t published.
  • Is the 87% token-reduction claim repeatable across diverse workflows, or specific to the benchmark TinyFish ran? No third-party verification yet.
  • How do free-tier rate limits behave under sustained agent traffic — per-account, per-key, or burst-windowed? The blog post specifies the per-minute caps but not the throttling shape.
  • The vendor’s positioning explicitly downgrades MCP relative to CLI + Skill. For wikis and ingest workflows that live inside Claude Code, is the MCP path adequate, or is the CLI install worth the extra setup for production runs?

Graph View

Backlinks