--- title: "Parallel Web Systems: The Web API Surface for Agents" description: "Parallel sells web search, typed decisions, research, discovery, and monitoring to AI agents through APIs, MCP servers, a CLI, and agent-harness plugins." canonical_url: "https://artificialcuriositylabs.ai/posts/2026-10-21-parallel-web-systems-the-web-api-surface-for-agents/" md_url: "https://artificialcuriositylabs.ai/posts/2026-10-21-parallel-web-systems-the-web-api-surface-for-agents.md" published_at: "2026-10-21T07:00:00.000Z" modified_at: "2026-10-27T07:00:00.000Z" tags: - "agents" - "mcp" - "search-api" --- Parallel is building web infrastructure for AI agents. Its index and APIs give agents sources to read, cited answers, researched records, candidate lists, and change alerts. The output an agent needs determines which part of the platform to call. ## Who Parallel is Parallel was founded by Parag Agrawal, Twitter's former CEO, [per TechCrunch](https://techcrunch.com/2026/04/29/parallel-web-systems-hits-2b-valuation-five-months-after-its-last-big-raise/). After a [$100M Series A](https://parallel.ai/blog/series-a) in November 2025, the company announced a [$100M Series B at a $2B valuation](https://parallel.ai/blog/series-b) in April 2026. ## What they are going after The company's [manifesto](https://parallel.ai/about) calls AIs "the web's second user" and argues for a "Programmatic Web" with verifiable provenance. [Index by Parallel](https://parallel.ai/blog/introducing-index-by-parallel) extends that argument to publishers: site owners can see how agents use their content, and participating publishers can be compensated. Parallel says it runs [its own web-scale index](https://parallel.ai/blog/parallel-search-api) beneath the product suite. Search returns excerpts; Extract reads known pages; Decisions answers closed questions; Responses gives cited answers; Task researches one record or report; FindAll builds a condition-checked list; and Monitor watches for changes. Each has a different unit of work and a different [pricing model](https://docs.parallel.ai/getting-started/pricing). ## The surface Where listed, prices are published rates; Chat has no published rate. Latencies are illustrative observations from small REST tests on one key between October 21 and 27, with different prompts and output sizes. The Image Search ranges cover five calls per mode; the Chat range covers five base calls, including one to the listed beta endpoint. Decisions reflects one two-question request. CLI timings appear in the client section. | Surface | What it does | Price | Measured REST latency | | ----------------------------------------------------- | ---------------------------------------------------------------------------- | --------------------------------------------------------------------------------- | ----------------------------------------------------------------- | | Search, `/v1/search` | Ranked excerpts; tiers turbo, fast, basic, advanced (default) | $0.001 turbo and fast; $0.005 basic and advanced | 0.5 s turbo; 0.75–1.4 s fast; 1.5–1.8 s basic; 1.3–3.0 s advanced | | Image Search, `/v1/images/search` | Image URLs with source pages; fast or advanced | $0.001 fast; $0.005 advanced | 0.7–1.0 s fast; 0.8–1.4 s advanced | | Extract, `/v1/extract` | Excerpts or full content for up to 20 URLs | $0.001 per URL | 0.6–7.8 s | | Decisions, `/v1/decisions` | Web-grounded probability, choice, or score answers to named questions | $0.001 per fast request, up to 100 related questions | 0.8 s for two questions | | Responses, `/v1/responses` | Cited answer; effort low, medium, high | $0.01, $0.05, $0.25 | 4.8 s low; 25.7 s high | | Task, `/v1/tasks/runs` | One researched record or report; nine processors, lite to ultra8x | $0.005 to $2.40 per run | 73 s on pro with a text schema | | Task Groups, `/v1/tasks/groups` | The same processors over many rows | $0.01 per row on base | 216 s for 10 rows | | Chat (beta), `/v1beta/chat/completions` | Chat completions; the model is a Task processor name | unpublished | 10.0–11.4 s on base | | Entity Search (beta), `/v1beta/findall/entity-search` | Ranked companies or people, no citations | $0.005 with 100 results | 1–3 s | | FindAll (beta), `/v1beta/findall/runs` | Cited candidates checked against stated conditions; preview, base, core, pro | $0.10 preview; $0.25 + $0.03 per match for base, up to $10 + $1 per match for pro | 7–11 min preview; about 9 min base | | Monitor, `/v1/monitors` | Watches the web; lite or base; event stream or snapshot | $0.003 lite, $0.01 base per check | ongoing | [Parallel's pricing page](https://docs.parallel.ai/getting-started/pricing) includes up to ten Search results in the base price and charges $0.001 for each additional result. Image Search includes up to 20 results. [FindAll is in beta](https://docs.parallel.ai/findall-api/findall-quickstart), so its endpoints and formats may change. Two examples make the product split concrete. [FindAll](https://parallel.ai/products/findall) can start without a list and look for San Francisco restaurants whose own sites publish vegetarian tasting menus and prices; each match still needs a source check. [Monitor](https://docs.parallel.ai/monitor-api/monitor-quickstart) can watch for changes in a retailer's published product price or stock status on a schedule. These are illustrative workflows, not live restaurant or inventory results. ## MCP, CLI, and plugins For new installs, [Parallel recommends one MCP server](https://docs.parallel.ai/integrations/mcp/quickstart) at `mcp.parallel.ai/mcp`. An anonymous connection on October 27 listed `web_search`, `web_fetch`, and `web_decisions`; the guide says authenticated clients also get `web_research` for cited answers. The unified guide lists no direct tool for FindAll, Monitor, Entity Search, or Image Search. The separate [Task MCP](https://docs.parallel.ai/integrations/mcp/task-mcp) still handles asynchronous research and list enrichment; a Docs MCP serves product documentation. My October 24 tests of the [standalone Search MCP](https://docs.parallel.ai/integrations/mcp/search-mcp) found that anonymous calls ignored mode, domain, and location overrides without a warning. Their excerpts averaged about 2,900 characters, but the response did not identify the search tier. Authenticated calls included a per-call cost field. I have not tested override handling on the unified server. The [CLI](https://docs.parallel.ai/integrations/cli), `parallel-cli` 0.9.3, covers search, extraction, research, FindAll, and monitoring. It sends `basic` search with a 60,000-character excerpt budget; REST defaults to `advanced`. In my CLI tests, `basic` took 1.3 seconds and `advanced` 4.1 seconds. The [Claude Code](https://docs.parallel.ai/integrations/claude-code-marketplace) and [Cursor](https://docs.parallel.ai/integrations/cursor-marketplace) plugins call the CLI, while the tested [OpenCode plugin](https://docs.parallel.ai/integrations/opencode-plugin) 1.2.0 fixed search at `advanced` and returned raw API JSON. Its `parallel-fetch` read a PDF that OpenCode's built-in webfetch refused for exceeding its 5 MB limit. ## Sharp edges REST and CLI responses returned usage counts without dollar costs, while authenticated standalone Search MCP responses included a per-call dollar cost. Extract can return HTTP 200 with failed URLs in `errors[]`, so a client that checks only the status code can miss a failure. A raw JSON schema sent to Task returned HTTP 422; wrapping it as `type: "json"` with `json_schema` started the run. Task runs and Task Groups take minutes and have no cancel route. `getStatus` on a nonexistent ID returns `unknown` with `is_active: true`, which can keep a naive poller running. Responses accepted `reasoning.effort: "bogus"`, ran at medium effort, and returned HTTP 200 instead of an error. When I used `auto`, the result did not reveal the effort tier it chose. Since the tier changes the price, I would set it explicitly before a paid call. ## So what Pick the endpoint by the output needed: Search and Extract for source pages, Responses for a cited answer, Decisions for a closed question, Task for a researched record, FindAll for a new list, and Monitor for changes. For cross-provider comparisons, see [the market map](https://artificialcuriositylabs.ai/posts/2026-09-12-search-for-agents-is-becoming-its-own-market/) and [the capability matrix](https://artificialcuriositylabs.ai/posts/2026-09-12-capability-matrix-for-agent-search-apis/). I’m using pinned `fast` Search for source discovery at $0.001 per request, then reading the source pages before using a result. The next test is whether Monitor catches changes to the source pages behind a FindAll candidate list.