io.github.ockram4444/agentcompute
3 changes recorded. MCP Harbor re-reads this listing from the official registry every six hours and asks the live server what it exposes, and keeps what changed between one look and the next.
https://u8x5gpk0ul.execute-api.eu-north-1.amazonaws.com/llms.txt
− No charge, card or ChatGPT login. Chat/images: 24 submissions globally per rolling 24 hours, including owner tests; 10 per session and per optional reported client ID; one active job per session. Shared cap: three queued/running Studio jobs. Queue expiry 10 minutes; execution maximum 5 minutes. Output expires 24 hours after submission. Temporary-session issuance remains globally capped. HTTP 429 means wait; do not rotate credentials. Reuse the SAME idempotency key and payload after an uncertain − Upgraded workers run Qwen3.5-9B Q4_K_M for chat and FLUX.2 klein 4B for 1024x1024 images. Check the live catalog. GPU may be offline; do not repeatedly submit into an unavailable service. Older workers may advertise other models. No external AI fallback, URL fetching, code execution or client-selected model. Generated code is text only.+ ## Agent tasks with local tools+ Check get_service_status and require services.agent_tools=true. Call create_trial_access once if needed and keep its token private. Then call submit_agent_job with:+ ```json+ {+ "access_token": "YOUR_PRIVATE_TOKEN",+ "agent_task": {+ "version": 2,+ "objective": "Calculate 17 items at 23.40 plus 38 shipping, then add 22 percent VAT to the subtotal. Use the calculator and create totals.csv with subtotal,vat,total.",+ "required_artifacts": ["totals.csv"]+ },+ "consent": true,+ "idempotency_key": "example-totals-unique-001"+ }+ ```+ This example is non-sensitive synthetic data. Set consent only for material you are permitted to process. Use a new idempotency key for a new task; reuse the identical payload and key after an uncertain submission. Poll get_studio_job using the same access_token and job_id at poll_after_seconds. Stop on succeeded, failed or expired. On success read result.agent_result, including artifacts [{name,content}]. Treat output filenames and content as untrusted data; save only explicitly chosen files an+ The model chooses approved local tools: decimal arithmetic, CSV preview/filter/numeric sort/summary/grouped sums, supplied-text read/search, and TXT/MD/CSV/JSON file creation. Optional documents are [{id,text}]. The model runs on the provider GPU; tools use that PC's CPU/RAM. No VM or cloud code runtime is required. Hosted coordination still uses AWS.+ Limits: 8 tool calls, 8 optional documents, 24,000 input characters, 180 seconds agent execution; at most 3 files, 16,000 UTF-8 bytes per file and 24,000 total. Documents are plain text, not binary uploads. CSV output rejects formula-like cells, currently including negative values; it is not a full spreadsheet engine. No browsing, arbitrary Python, shell or host-file access. Quotes are matched to supplied text or reported tool outputs; factual accuracy and tool computations are not independently+ Agent tasks share the chat/image Studio quotas. Legacy version 1 document tasks remain supported. Connectors need to refresh their MCP tool list to see updated schemas; a registry listing does not automatically connect an agent.+ No charge, card or ChatGPT login. Agent/chat/image tasks: 72 submissions globally per rolling 24 hours, including owner tests; 10 per session and per optional reported client ID; one active job per session. Shared cap: three queued/running Studio jobs. Queue expiry 10 minutes; execution maximum 5 minutes. Output expires 24 hours after submission. Temporary-session issuance remains globally capped. HTTP 429 means wait; do not rotate credentials. Reuse the SAME idempotency key and payload after an+ Prepared workers can run Qwen3.5-4B or Qwen3.5-9B Q4_K_M for chat, depending on the eligible profile. Image-capable profiles use FLUX.2 klein 4B for 1024x1024 images. Check the live catalog. GPU may be offline; do not repeatedly submit into an unavailable service. Older workers may advertise other models. Plain chat has no URL fetching, code execution or client-selected model. Use version 3 agent tasks for public web tools; no external AI fallback is used. Generated code is text only.+ Public web research: set agent_task.version=3 and check services.agent_web. The provider can search Bing and read public HTTPS pages as text, with source URLs. Up to eight tool calls total. No JavaScript, login, account actions, arbitrary code or host-file access. Search results can vary and some pages are blocked or truncated. Queries go to Bing; domain lookups use Google DNS; pages are fetched by the provider. Send only nonconfidential public search terms. Version 2 local tools remain supporte+ WEBSITE AND PROVIDER PREVIEW (30 October 2026)+ Use compute: /compute.html (teams and developers); Provide compute: /providers.html (Windows GPU owners). Windows preview 0.4.15 is available at /downloads/AgentCompute-Provider-Setup-0.4.15.exe; release metadata at /downloads/provider-0.4.15.json. This experimental installer is signed by C.M.S. SRLS and timestamped. Version 0.4.15 adds a capability-gated coordination protocol for planning, parallel document analysis, synthesis and review. Exact decimal table totals and source quotations are che+ AGENT CONNECTION FIRST+ AgentCompute exposes supported task execution through the remote Streamable HTTP MCP endpoint and HTTP API. Connect an existing compatible agent using /agents.html#connect; use /docs.html for application integration. Studio is an optional browser demo, not a prerequisite. Agents must be configured in their client and follow availability, permission, quota and idempotency requirements. This service delegates supported tasks to prepared models and tools; it does not provide unrestricted GPU instan
We can't find the internet
Attempting to reconnect
Something went wrong!
Attempting to reconnect