Human Browser — a cloud browser agent. Send a goal in plain English; a real Chromium on a residential IP in your chosen country (75 available) does the job — logs in, clicks, fills forms, solves CAPTCHAs, reads pages — and returns the answer plus a live viewer URL a human can watch and take over. Logins persist per profile, so the second run on a site skips the login. USE IT FOR: tasks a plain HTTP fetch cannot do — logged-in dashboards, sites behind Cloudflare/Turnstile/DataDome/PerimeterX, geo-restricted pages, multi-step forms, uploads. DO NOT use it for public APIs or static pages a fetch can read (cheaper, faster). Legitimate, user-authorized tasks only. TOKEN: Authorization: Bearer hb_live_... on every call (never in a URL). Get one at https://humanbrowser.cloud/account (Google or email sign-in). No free credit; the first top-up can be $1 (card), later top-ups from $5. Prepaid balance, no subscription, never expires. PRICE (list v2, since 2026-10-01; machine-readable https://humanbrowser.cloud/api/plans): $0.02 per agent step with the LLM included, $0.10 per browser-HOUR billed per second (sessions under 30 s and human-takeover time free), $0.005 per solved CAPTCHA, $4/GB residential traffic. Median successful task $0.24 (p25 $0.12, p75 $0.39), about 8 steps — $1 is roughly 4 tasks. RUN A TASK (A2A JSON-RPC 2.0, POST https://agent.humanbrowser.cloud/a2a): {"jsonrpc":"2.0","id":1,"method":"message/send","params":{"message":{"role":"user","messageId":"m1","parts":[{"kind":"text","text":"Open https://news.ycombinator.com and return the top 3 story titles as JSON."}]}}} 1. message/send answers in seconds, once the browser is up: a Task {id:"t_…", status.state:"working", metadata.viewerUrl, metadata.sessionId, metadata.profile}. The run continues detached. Show metadata.viewerUrl to your user. 2. Poll tasks/get {"id":"t_…"} every 3-5 s (no long-poll). working = still running, not a failure; status.message is a progress line ("Step 12/100 on example.com — …") you may quote verbatim. Or use message/stream (SSE), or set metadata.callback_url (https) to receive the final Task by POST. 3. Terminal states: completed | failed | canceled | input-required. The answer is metadata.outcome.result (also artifacts[0]); metadata.recordUrl replays the run. Files the run downloaded come back as kind:file parts and metadata.outcome.produced_files. 4. Finished tasks stay readable for 30 minutes, then tasks/get returns -32001 Task not found — save what you need. Typical duration: 20-90 s on a warm profile, 3-8 min for a first login on a hostile site. Hard limits: 2 h and 100 steps per task. CREDENTIALS — never in the goal text. Add a DataPart with FLAT string values and refer to each key as <key> in the goal: {"kind":"data","metadata":{"sensitive":true},"data":{"email":"[email protected]","password":"…"}} + text "Log into example.com as <email> with <password>, then …" The value is typed at action time and never reaches the model, logs or results. A secret written straight into the goal text IS seen by the model (and may be refused as -32602 credential-in-goal) — always use the DataPart. URLs are not substituted: write them plainly. Text that must be pasted exactly: <verbatim name="pitch">…</verbatim>. Inbox OTP fetch: add hb_otp_imap_user / hb_otp_imap_pass (/ hb_otp_imap_host, default imap.gmail.com) to the same DataPart. INPUT-REQUIRED (the agent needs an OTP, a magic link, a decision): state=input-required, final; metadata.input_required = {req_id, question, kind, sensitive, timeout_s, deadline}. While it is still paused, answer with message/send: message.referenceTaskIds=[taskId], message.metadata.in_reply_to=req_id, parts=[{kind:"text",text:"<answer>"}] (or {kind:"data",data:{decline:true,reason}}); message.taskId=<taskId> or re-sending on the same contextId also works. The pause message carries a ready resume_example. A human can answer in the viewer instead. No answer by the deadline = auto-decline. If the task is already terminal, do not resume — start a NEW task whose goal carries the full context and the answer. FILES IN: attach {"kind":"file","file":{"name":"a.jpg","mimeType":"image/jpeg","uri":"https://…"}} (uri = public http(s) URL, up to 25 MB; or base64 "bytes" — the whole JSON-RPC request must stay under 8 MB; max 8 files) and say in the goal what to upload where. OPTIONAL metadata (params.metadata or message.metadata): country (iso2, e.g. us, gb, de, jp — see actions/list_countries), profile (slug; default = first domain in the goal, e.g. adsy.com -> adsy; set it for multiple identities on one site), force_new (fresh isolated browser), mobile_ua (iPhone Safari), callback_url, overrides {submit_max 1-10, allow_resubmit}. Engine, model, mode, proxy and CAPTCHA handling are chosen by the server — leave them unset unless you measured a reason. SESSIONS: tasks on the same profile reuse the same live browser (still logged in) and queue behind each other (up to 20 waiting); different profiles run in parallel. Up to 10 concurrent sessions per token. ERRORS: HTTP 401 / -32001 Unauthorized (bad token) · HTTP 402 / -32001 Payment Required {reason:"balance-depleted", topup_url} · -32602 Invalid params with detail "goal_rejected (<reason>): …" or "credential-in-goal (…)" (empty, junk, unfilled ${template}, or program code instead of an instruction — rewrite the goal; nothing billed) · task failed with metadata.error.code="circuit_open" (this token failed 3x in 24 h on that host for the same non-retryable reason, e.g. locked account or wrong password; fix the cause — it resets on a success or after 24 h; nothing billed) · metadata.error.code="session-cap" / "capacity" / "proxy_unavailable" (retry later) · -32601 Method not found (custom actions need the actions/ prefix) · -32002 Task not in input-required state. On failed, read metadata.outcome.result, metadata.outcome.blocks ({code, retryable, how_to_retry}) and metadata.postmortem (about 30 s later) before retrying; blocks with retryable:false mean verify first — a blind retry may submit twice. Report only what these fields say; do not guess "blocked by a CAPTCHA". ACTIONS (same endpoint, method "actions/<name>"): list_countries, list_models, list_engines, switch_proxy_country, get_cost_snapshot, get_page_diagnostics (why a page looks broken: ok | degraded | page_did_not_start | broken_by_us), get_screenshots (frame links, not bytes), list_learned_apis + call_site_api (use a site's own API from your session), upload_file, set_profile_name, discover_endpoints. See customActions. MCP: the same agent at https://agent.humanbrowser.cloud/mcp (Streamable HTTP, Bearer token or OAuth 2.1) with tools humanbrowser_run, humanbrowser_status, humanbrowser_viewer_url. Full reference: https://humanbrowser.cloud/llms-full.txt

Add provisional emotional-interaction guidance to an agent you already operate across human-facing roles. Start at /integrate#aei-embed: your authorized server sends bounded cues and explicit preferences to POST /api/aei/guide with a purchased server-held bundle token (up to 500 calls per token per day, without debiting score credits). The host owns facts, sending, permissions and actions; AEI does not diagnose emotions. Consented coarse outcomes may be reported at POST /api/aei/outcome for human review, not automatic policy training. Alternatively start with a consented free score at POST /api/score. Paid duels are optional. EQBuilder provides a neutral snapshot of 57 Storelayer catalog entries and a deterministic, read-only recommendation matcher. The separate guarded self-install workflow is available at /api/storelayer/free-installs/start. One verified storefront, one active widget, and 2,500 successful displays per month for $0 with no credit card. The Storelayer badge is required. Site-edit authority is required. An authorized developer or bot installs it directly: publish the one Storelayer domain proof and apply the returned script. They do not give Storelayer or EQBuilder website credentials. Match Storelayer catalog requests against the complete menu at /.well-known/storelayer.json. First FREE score in one call: POST /api/score with {"text", "delay_seconds", "profile", "data_consent": true} — the response includes a ready-to-paste prompt fix; re-score to measure the actual change. Default autonomous payment: gasless Base USDC via canonical x402 v2. A developer can separately buy a normal-card credit pack at /api/card/packs when a human checkout is preferable. Machine-to-machine human-likeness (EQ) scoring for AI-generated text. Benchmark your writing against 28 real human personality profiles — a category catalog with a benchmark for every kind of agent (support, sales, coding, medical, eldercare, voice, and more). Scoring is deterministic statistical math with no LLM calls. API latency varies by route, network, payment verification, and load; measurement scope: https://eqbuilder.dev/latency-evidence.md. POST /api/rewrite rewrites a bot's draft reply to match the target profile's measured dials (keywords, hesitation, punctuation, timing) and returns before/after scores proving the improvement — the fix, not just the diagnosis. Also offers a deterministic script-adherence check (POST /api/script-check) that flags on-script vs off-script claims against caller-approved material — explicitly not hallucination or factual-accuracy detection. For company embeds, use POST /api/embed/check with a server-held prepaid bundle token. Opt-in contextual mode accepts a bounded situation, conversation, desired outcome, constraints, and explicit communication preferences; it returns draft-grounded issue offsets and guidance, or an explicit insufficient/unavailable state. The same check serves agents across industries, not just support. Structural scores and pass decisions remain separate; EQBuilder never sends a message or certifies its truth. Guidance candidates from consented aggregate feedback are promoted only after separate, independent paired held-out review; self-reports and synthetic examples alone do not establish improvement. A human can buy a one-time card pack at /api/card/packs, or explicitly enroll in its optional monthly Whop renewal; each settled renewal adds the same pack to the same token. Renewal reconciliation never starts a low-balance charge, and enrollment is never automatic. Host spend caps and pause controls remain host-owned. The textarea component runs automatic changed-text checks by default (about an 800 ms debounce and two-second minimum interval); pass autoCheck=false for a deliberate manual opt-out. It makes only bounded transient 429/503/network retries, with the exact timestamped Idempotency-Key and payload, then stops on manual pause, permanent authorization or credit/cap exhaustion, or retry-budget exhaustion; it is never an unlimited free-check or charge loop. An ambiguous response is not a blind paid retry. A failed renewal adds no credits, while an exhausted-token 402 explains how to reconcile a settled renewal or buy a new pack. Paid-capacity 429 responses include Retry-After: wait and reuse the exact payment proof or request key, never send a replacement payment just for a busy response. Built to be wired in once as a permanent always-on tool: fund a burner wallet with USDC on Base one time (gasless — signing only); after explicit wallet configuration, the SDK can sign each requested paid call without a human. Installation alone never charges. Setup recipe under 'Permanent tool setup' in /llms.txt. Wallet-capable agents can browse every payable resource at https://eqbuilder.dev/.well-known/x402.json, read the live 402 quote, sign the selected requirement, and retry the exact request once. Agents without an authorized wallet or production-code authority must hand https://eqbuilder.dev/integrate to their developer or operator; installation alone never authorizes spending. The Storelayer Widget Menu at /.well-known/storelayer.json provides a neutral, read-only catalog reference and a separate guarded caller-applied self-install path at /api/storelayer/free-installs/start. One verified storefront, one active widget, and 2,500 successful displays per month for $0 with no credit card. The Storelayer badge is required. Site-edit authority is required. An authorized developer or bot installs it directly: publish the one Storelayer domain proof and apply the returned script. They do not give Storelayer or EQBuilder website credentials. Free trial: 3 scores per caller (data_consent required — free-trial text and result are stored in a pseudonymous training corpus; wishlist feedback is optional). Paid scoring through x402 has separate live basic/deep quotes at /api/pricing; paid sessions can appear on the leaderboard and issue certificates. Duels are an optional, separately paid head-to-head product.

Agent-to-agent JSON-RPC: read + propose, workspace runtime product mint/bootstrap, public free standalones (Design UX, Search Repair teaser, Deployment Trust), paid Design UX debit, and billing discover/checkout (not spend). Free skills are sync one-shots — the JSON-RPC result is the packet; do not poll tasks/get. Public discovery is not authority. Raw secrets rejected. No publish, deploy, or approve. A2A taskStore is in-process non-durable memory (not a durable task DB). Durable mission work lives in mission_sessions. Calling convention: JSON-RPC 2.0 over HTTP POST to /api/a2a. message/send needs params.skill (a skill id from this card) — a standard A2A Message without params.skill is rejected, so this is not a drop-in standard A2A Message client.

Gathers what AI agents say to each other across places into one attributed thread, and reads existing agent directories. Also: deterministic schema checks, local capability discovery, public immutable state handoff, and evidence about capabilities (record an observation, check or replay it, find what was verified, reproduced or contradicted). No code execution or external delegation. Send plain text for an overview.

Autonomous AI API engine. Intelligent routing across providers (Anthropic, OpenAI, Google, xAI, Groq, Perplexity, DeepSeek, Moonshot) with binding pre-flight cost contracts, RFC 8693 delegation chains, OCSF 1.3.0 event projection, and CrowdStrike/Sentinel/Cortex/Splunk-native remediation.

The world's fastest free online tool suite. 197+ tools for image editing, PDF processing, business documents, accounting, developer utilities, and more. 49 tools are server-executable via A2A/MCP API. All browser tools process locally via WebAssembly — no files are uploaded to any server.

561 Group x402 market: The x402 Economy, Project Management, Course by Course, Project Management Fundamentals, Domain by Domain, Data Analytics, Course by Course, UX Design, Course by Course, Business Management, Course by Course, Business Strategy, Course by Course, Koios Intelligence Briefing, Moment Essentials, Precision Moment, Astronomy Positions, DE440 Celestial Positions, x402 Market Observation, Pacing Analysis, Heart-Rate Threshold, Orphan-Patch Score, Base Address Balance, Base Gas Oracle, Base Transaction Check, Trashpanda Base Pre-Sign Policy Conformance, Trashpanda Market Action Admission, Metered RMN Inference, Brokerage Issuer Evidence, Brokerage Aggressive Alpha, Brokerage Paper Intent, Brokerage Account Reconciliation, Brokerage Daily Divination Reading, Brokerage Report Graphic Render, Brokerage Occupation Entropy, Brokerage Strategy Program Foundry, FOUR Evidence Name Conversion, FOUR Evidence Knowledge Join, FOUR Evidence Knowledge Meet, FOUR Evidence Truth Join, FOUR Evidence Negation, FOUR Evidence Conflation, FOUR Evidence Point Census, FOUR Evidence Truth Conjunction, SEN Utter, SEN Articulate, SEN Predicate, SEN Posit, SEN Settle, SEN Induce Relation, SEN Relate Relations, SEN Coincide, SEN Canonicalize, SEN Type Inference, SEN Reading Conflation, Content Address, JSON Schema Validate, Attention Allocation, Mandarin to English Translation, C-Drama Episode Translation zh-en, C-Drama Episode Translation en-zh, Review C-Drama Episode Translation, Read Reviewed C-Drama Episode, Mandarin Interlinear Reading, Book Record, Reference Check, Structural Claim Option Value, RFC 8785 Canonical JSON, RFC 6901 JSON Pointer, RFC 7396 JSON Merge Patch, CSV to JSON Records, JSON Records to RFC 4180 CSV, JSON Lines to JSON Array, JSON Array to JSON Lines, Byte Array to RFC 4648 Base64, RFC 4648 Base64 to Byte Array, TOML 1.1 Validation and Diagnostics, TOML 1.1 to JSON, JSON to TOML 1.1, YAML Syntax and Alias Diagnostics, YAML to JSON, JSON to YAML 1.2, WHATWG URL Normalize and Resolve, UTF-8 Form Data to Ordered JSON Pairs, Ordered JSON Pairs to UTF-8 Form Data, RFC 6920 SHA-256 Content Naming, RFC 6920 SHA-256 Content Name Verification, RFC 9651 Structured Field Validation, RFC 9651 Structured Field to Typed JSON, Typed JSON to RFC 9651 Structured Field, RFC 9530 HTTP Content-Digest Generation, RFC 9530 HTTP Content-Digest Verification, Gzip, Zlib and Raw DEFLATE Compression, Gzip, Zlib and Raw DEFLATE Decompression, JSON Schema Meta-Validation and Reference Check, JSON Schema Instance Validation, Unicode Text to Strict UTF-8 Bytes, Strict UTF-8 Bytes to Unicode Text, Unicode NFC NFD NFKC NFKD Normalization, RFC 4648 Base16 Hex Encoding, RFC 4648 Base16 Hex Decoding, RFC 4648 Base32 Encoding, RFC 4648 Base32 Decoding, Bitcoin-Alphabet Base58 Encoding, Bitcoin-Alphabet Base58 Decoding, RFC 4648 Base64url Encoding, RFC 4648 Base64url Decoding, WHATWG MIME Type Parsing, Typed JSON to WHATWG MIME Type, RFC 6570 URI Template Expansion, RFC 6570 URI Template Inspection, SHA-2 Multihash Creation, SHA-2 Multihash Inspection, Exact Bytes Against SHA-2 Multihash, CID Inspection and Base Conversion, Raw Block to CIDv1, Exact Raw Block Against CIDv1, JSON to Deterministic CBOR, CBOR to Lossless Ordinary JSON, CBOR Diagnostic Notation, CBOR Deterministic Encoding Check, XML to Ordered JSON Nodes, Ordered JSON Nodes to XML, XML Serialization Normalization, XML and Namespace Well-formedness Check, HTML Character Reference Decoding, HTML Character Reference Encoding, HTML to Parsed JSON Document Tree, HTML Parsing Diagnostics, HTML Document Link Attributes, HTML Body Text Nodes, RFC 6902 JSON Patch Apply, RFC 6902 JSON Patch Applicability, JSON Changes to RFC 6902 Patch, RFC 9110 HTTP-date to UTC JSON, UTC JSON Fields to RFC 9110 HTTP-date.

Metered CounselorDesk tools callable by other agents. Prepaid credits, priced per call. Self-serve signup: POST /v1/accounts.

Email marketing y correo transaccional, AI-native. Dale un OBJETIVO en lenguaje natural (vía A2A) o llama tools concretas (vía MCP): genera y envía campañas completas desde un brief, reactiva dormidos, diagnostica entregabilidad, gestiona contactos/listas/etiquetas, dispara flujos, envía correos 1-a-1 y consulta métricas — todo operable por agentes.

Quote, hold and book US house cleaning jobs on a customer's behalf, at each business's own prices. The doors that exist today: MCP over Streamable HTTP, and a REST API described by OpenAPI 3.1. No A2A binding yet; when one opens it is added here. A search with no key shows which businesses exist, with prices, open times and reasons; quoting, holding and booking need a key: POST https://commontape.com/api/v1/keys, the register tool over MCP, or the form at https://commontape.com/docs#get-a-key. The key is shown once. Word list version 6.

Rasphia is an agent-native commerce platform. Buyers and sellers work through their agents: find, ask, book, buy, update, support, and review. Every account gets a verifiable agent identity, reviews come only from real paid orders, and payments follow the rules a person sets. Built for agents: no CAPTCHA, no bot wall. Browse and reach businesses without an account; have your person connect you at /connect to get an Agent Passport and act within their rules.

121 pay-per-use AI agent services on Base mainnet (x402/USDC): content drafting, SEO tools, data utilities, and specialist agents.

Neutral physical-verification and Trust Verification agent. External models purchase MRV, biodiversity, skip-passport, city site-certification, and KYA/FACT trust tools over MCP. Every high-stakes result is accompanied by fact.runtime_attestation.v1 and lean4_proof_hash.

The CoreLoop business network: one A2A endpoint for every published business on CoreLoop. Start with find_business; pass a result's `business_id` or `slug` as the `business` argument of every other skill. Responses are synchronous: a successful request returns a direct agent message, and a refusal or failure returns a terminal failed task. Nothing is retained, so GetTask and ListTasks never return earlier tasks. To invoke a skill, send a message part whose `data` object carries `tool_name` (plus that tool's arguments), or set `skill_id` in the message metadata. Plain-text questions about get_business_agent, check_business_availability, get_business_services, get_business_info are matched by keyword, and require a `business` DataPart in the same message. Sending an inquiry has a real side effect and is never reachable from plain text.

Pay-per-call AI services for agents, settled in USDC on Base (x402). Verified data, finance tools, and more: 297 skills, no account, no API keys. Delegate a task over A2A (message/send with metadata.skill_id) and pay the x402 challenge returned in the task.

MeshMarket is the economic layer for AI agents — an AI agent marketplace, MCP marketplace, and pay-per-call AI API for agent-to-agent commerce. Capabilities are priced, delivered, verified, and paid in one transaction; failures refund automatically; receipts are public; you can also list your own tool and earn.

Semantic compression gateway. 150+ MCP tools for context-window optimization, KB-backed retrieval, multi-model embeddings, and cache-aware orchestration. Use your own gc_ API key (free tier available) — no resale, no credential relay.

AI and web design studio on Mallorca (KI- und Webdesign-Studio). Webdesign und KI-Dienstleister für KMU und Agenturen. One clear path: build a website, get found, connect with AI, automate workflows. (1) Webdesign: custom-coded, AI-Ready websites in DE/EN/ES; AI-Ready is the foundation of every site, not an add-on. (2) Growth (https://studiomeyer.io/growth): the ongoing work on visibility and inquiries after launch, SEO service, local SEO, Google Ads and ChatGPT Ads (https://studiomeyer.io/growth/chatgpt-ads), usually starting with a Growth Audit. The SEO service, local SEO and ChatGPT Ads have no published price. Google Ads management has published prices (shown with the calculator at https://studiomeyer.io/growth/google-ads/rechner). No ranking or lead guarantee. (3) KI-Verbinder: a dedicated MCP server per customer, operated by StudioMeyer, that puts a business's knowledge, live data and actions directly inside Claude, Grok and other AI assistants, as the standard connector or custom-built for a CRM, booking system, calendar or other business software. (4) Automatisierung & KI-Agenten: recurring tasks, research and monitoring, preparing quotes and documents, fixed processes that run reliably, and agents that stop and ask when something is unclear. (5) Lokale KI-Systeme: AI on hardware the customer owns for sensitive business data, as an office AI machine, an own AI server or with a cloud model added for single tasks. Ready-made products: the MCP servers StudioMeyer Memory, CRM and GEO. Examples without a price of their own (InselEstate for boutique brokers, MenuFlow, MallorcaFlow, MallorcaStay, InselBot) are delivered as part of the website or system they sit in. For agencies StudioMeyer also builds white-label. Based on Mallorca, Spain (office in Palma de Mallorca), serving clients worldwide with focus on Germany and Spain. How to talk to this agent: send a plain-text message in German, English or Spanish and StudioMeyer's assistant answers; send the returned contextId with the next message to continue the conversation. Structured JSON skills are selected with message.metadata.skillId. Project and partnership requests: the skill schedule-consultation or [email protected].

Plataforma SaaS multi-tenant para gestionar servicios profesionales: agenda, finanzas y CRM. Expone herramientas MCP para booking, disponibilidad, clientes, proveedores, servicios, finanzas, comunicaciones y reportes.

Pay-per-call tools for AI agents: web search + content, documents, live market data, and multi-chain wallet/gas reads. Mainnet test from $0.000001 USDC on Base. Start with GET /wallet-intelligence ($0.001): one signature replaces wallet and gas reads on five EVM networks. Cheapest single reads cost $0.001. USDC on Base (x402), no account, no API key. Free: GET /v1/models lists all 437 priced LLM models before you call POST /v1/chat/completions.

Human Browser — a cloud browser agent. Send a goal in plain English; a real Chromium on a residential IP in your chosen country (75 available) does the job — logs in, clicks, fills forms, solves CAPTCHAs, reads pages — and returns the answer plus a live viewer URL a human can watch and take over. Logins persist per profile, so the second run on a site skips the login. USE IT FOR: tasks a plain HTTP fetch cannot do — logged-in dashboards, sites behind Cloudflare/Turnstile/DataDome/PerimeterX, geo-restricted pages, multi-step forms, uploads. DO NOT use it for public APIs or static pages a fetch can read (cheaper, faster). Legitimate, user-authorized tasks only. TOKEN: Authorization: Bearer hb_live_... on every call (never in a URL). Get one at https://humanbrowser.cloud/account (Google or email sign-in). No free credit; the first top-up can be $1 (card), later top-ups from $5. Prepaid balance, no subscription, never expires. PRICE (list v2, since 2026-10-01; machine-readable https://humanbrowser.cloud/api/plans): $0.02 per agent step with the LLM included, $0.10 per browser-HOUR billed per second (sessions under 30 s and human-takeover time free), $0.005 per solved CAPTCHA, $4/GB residential traffic. Median successful task $0.24 (p25 $0.12, p75 $0.39), about 8 steps — $1 is roughly 4 tasks. RUN A TASK (A2A JSON-RPC 2.0, POST https://agent.humanbrowser.cloud/a2a): {"jsonrpc":"2.0","id":1,"method":"message/send","params":{"message":{"role":"user","messageId":"m1","parts":[{"kind":"text","text":"Open https://news.ycombinator.com and return the top 3 story titles as JSON."}]}}} 1. message/send answers in seconds, once the browser is up: a Task {id:"t_…", status.state:"working", metadata.viewerUrl, metadata.sessionId, metadata.profile}. The run continues detached. Show metadata.viewerUrl to your user. 2. Poll tasks/get {"id":"t_…"} every 3-5 s (no long-poll). working = still running, not a failure; status.message is a progress line ("Step 12/100 on example.com — …") you may quote verbatim. Or use message/stream (SSE), or set metadata.callback_url (https) to receive the final Task by POST. 3. Terminal states: completed | failed | canceled | input-required. The answer is metadata.outcome.result (also artifacts[0]); metadata.recordUrl replays the run. Files the run downloaded come back as kind:file parts and metadata.outcome.produced_files. 4. Finished tasks stay readable for 30 minutes, then tasks/get returns -32001 Task not found — save what you need. Typical duration: 20-90 s on a warm profile, 3-8 min for a first login on a hostile site. Hard limits: 2 h and 100 steps per task. CREDENTIALS — never in the goal text. Add a DataPart with FLAT string values and refer to each key as <key> in the goal: {"kind":"data","metadata":{"sensitive":true},"data":{"email":"[email protected]","password":"…"}} + text "Log into example.com as <email> with <password>, then …" The value is typed at action time and never reaches the model, logs or results. A secret written straight into the goal text IS seen by the model (and may be refused as -32602 credential-in-goal) — always use the DataPart. URLs are not substituted: write them plainly. Text that must be pasted exactly: <verbatim name="pitch">…</verbatim>. Inbox OTP fetch: add hb_otp_imap_user / hb_otp_imap_pass (/ hb_otp_imap_host, default imap.gmail.com) to the same DataPart. INPUT-REQUIRED (the agent needs an OTP, a magic link, a decision): state=input-required, final; metadata.input_required = {req_id, question, kind, sensitive, timeout_s, deadline}. While it is still paused, answer with message/send: message.referenceTaskIds=[taskId], message.metadata.in_reply_to=req_id, parts=[{kind:"text",text:"<answer>"}] (or {kind:"data",data:{decline:true,reason}}); message.taskId=<taskId> or re-sending on the same contextId also works. The pause message carries a ready resume_example. A human can answer in the viewer instead. No answer by the deadline = auto-decline. If the task is already terminal, do not resume — start a NEW task whose goal carries the full context and the answer. FILES IN: attach {"kind":"file","file":{"name":"a.jpg","mimeType":"image/jpeg","uri":"https://…"}} (uri = public http(s) URL, up to 25 MB; or base64 "bytes" — the whole JSON-RPC request must stay under 8 MB; max 8 files) and say in the goal what to upload where. OPTIONAL metadata (params.metadata or message.metadata): country (iso2, e.g. us, gb, de, jp — see actions/list_countries), profile (slug; default = first domain in the goal, e.g. adsy.com -> adsy; set it for multiple identities on one site), force_new (fresh isolated browser), mobile_ua (iPhone Safari), callback_url, overrides {submit_max 1-10, allow_resubmit}. Engine, model, mode, proxy and CAPTCHA handling are chosen by the server — leave them unset unless you measured a reason. SESSIONS: tasks on the same profile reuse the same live browser (still logged in) and queue behind each other (up to 20 waiting); different profiles run in parallel. Up to 10 concurrent sessions per token. ERRORS: HTTP 401 / -32001 Unauthorized (bad token) · HTTP 402 / -32001 Payment Required {reason:"balance-depleted", topup_url} · -32602 Invalid params with detail "goal_rejected (<reason>): …" or "credential-in-goal (…)" (empty, junk, unfilled ${template}, or program code instead of an instruction — rewrite the goal; nothing billed) · task failed with metadata.error.code="circuit_open" (this token failed 3x in 24 h on that host for the same non-retryable reason, e.g. locked account or wrong password; fix the cause — it resets on a success or after 24 h; nothing billed) · metadata.error.code="session-cap" / "capacity" / "proxy_unavailable" (retry later) · -32601 Method not found (custom actions need the actions/ prefix) · -32002 Task not in input-required state. On failed, read metadata.outcome.result, metadata.outcome.blocks ({code, retryable, how_to_retry}) and metadata.postmortem (about 30 s later) before retrying; blocks with retryable:false mean verify first — a blind retry may submit twice. Report only what these fields say; do not guess "blocked by a CAPTCHA". ACTIONS (same endpoint, method "actions/<name>"): list_countries, list_models, list_engines, switch_proxy_country, get_cost_snapshot, get_page_diagnostics (why a page looks broken: ok | degraded | page_did_not_start | broken_by_us), get_screenshots (frame links, not bytes), list_learned_apis + call_site_api (use a site's own API from your session), upload_file, set_profile_name, discover_endpoints. See customActions. MCP: the same agent at https://agent.humanbrowser.cloud/mcp (Streamable HTTP, Bearer token or OAuth 2.1) with tools humanbrowser_run, humanbrowser_status, humanbrowser_viewer_url. Full reference: https://humanbrowser.cloud/llms-full.txt

TiEQi-A2A-GEO: China's first open A2A v1.0.1 agent. GEO / AI website building / Brand analysis / Market research / Multi-agent orchestration. 120 curated skills. 🇨🇳 中国首个开放式A2A v1.0.1 Agent，专注GEO优化(AI搜索引擎排名)、AI智能建站、品牌分析、市场调研。