AIsa

Agents act in outside systems through one key. The catalog puts thousands of APIs behind a single AIsa key, covering search, finance, social, company and people data and SaaS tools, served over REST and an MCP server. The server's five tools (list_categories, search, get_details, use and batch_use) reach 950+ APIs, with write actions such as Twitter Autopilot's approved actions and the Gmail and Notion Skills; the docs note that writes may need OAuth or an account connection for each capability.

SourceAIsa, aisa.one/llms.txt (MCP), aisa.one/docs llms-full.txt (Architecture, Security Evaluation Guide) and the homepageread 2026-09-25

AIsa runs no workflow. A Skill packages instructions, tool selection and workflow steps that the customer's own agent follows, and the docs frame the choice between a Skill and the direct API as a question of who owns orchestration: the agent or the application, not AIsa. The batch_use tool runs several calls in one request, which is batching, not sequencing. The Foundry card offers preconfigured OpenClaw instances, an open source runtime AIsa hosts but did not build, documented only as a card.

Sourceaisa.one/docs/llms-full.txtread 2026-09-25

Nothing retrieves over the customer's own knowledge. The search, crawl and research APIs AIsa resells, such as Tavily, Exa and Perplexity, query the public web, not the customer's corpus, and the Jina embedding and rerank endpoints are building blocks a customer would use to make an index, not an index AIsa maintains.

SourceAIsa, aisa.one/docs llms-full.txt (Search APIs, Create Embeddings, Rerank Documents)read 2026-09-25

Spend limits are enforced, and approval is left to the agent's host. The MCP server's get_details returns what an operation costs and a suggested maximum, and use and batch_use take max_price_usd and refuse anything above it before any charge. Each API key carries its own spending quota, capped at the wallet balance. The docs set an approval rule (present scope and cap, get explicit approval, never run a paid call in the planning turn), but that rule is instructions the customer's agent follows, not an approval surface AIsa ships. The spend caps are the customer's to set.

SourceAIsa, aisa.one/llms.txt (MCP) and aisa.one/docs llms-full.txt (Authentication, Machine Payments for AI Agents)read 2026-09-25

Access controls are thin and there is no compliance posture. Keys are scoped by their own spending quotas and meant to be used one per service or environment. Dashboard sign in supports Google and GitHub OAuth, and the docs advise SSO so revocation follows the identity provider; MCP clients sign in with OAuth tokens that can each be revoked. No roles or organization permissions are documented, and no attestation or certification appears in the docs or on the site. The Enterprise card lists zero data retention, a custom MSA and DPA, and advanced access controls without detail.

SourceAIsa, aisa.one/docs llms-full.txt (Authentication, Security Evaluation Guide) and aisa.one/pricingread 2026-09-25

Each request is logged, but there is no trace of a run. Console Usage and Logs shows summary cards, a usage trend and a split by model, plus a table per request showing the model or API, cost, cache hits and status, filterable by key. Each MCP use call reports what it actually spent. That shows what was called and what it cost, one call at a time; no view ties calls into an agent run or shows the steps and reasoning between them.

SourceAIsa, aisa.one/docs llms-full.txt (Usage Logs, Authentication) and aisa.one/llms.txt (MCP)read 2026-09-25

AIsa sits in the request path, from the agent to an AIsa endpoint to the upstream capability and back, with no memory layer. The docs leave state handling to the application, and the usage log is a record of activity, not memory.

SourceAIsa, aisa.one/docs llms-full.txt (Architecture and Integration Boundaries)read 2026-09-25

A hosted gateway runs at api.aisa.one. The Enterprise card offers dedicated compute and custom infrastructure built around the customer's deployment model and compliance requirements, available on request, but names no region list, customer environment option or way to select one.

SourceAIsa, aisa.one/pricing (Enterprise) and aisa.one/docs llms-full.txt (Architecture)read 2026-09-25

The Agent Card and the Skills catalog list 40+ installable Skills, each a complete reusable outcome that does its own job when selected, such as SEO Keyword Research, KOL Creator Discovery, MarketPulse, Twitter Autopilot and Prediction Market Arbitrage. AIsa also ships ready made agents, such as AIsa CIO, an investment agent, and AIsa Labs templates for marketing, research and SEO. The Skills library arrived with the Bundled Skills System on 21 July.

Sourceaisa.one/.well-known/agent-card.jsonread 2026-09-25

Only a request from an agent or application starts work, since AIsa is a gateway that callers invoke. No schedule, webhook or event source starts anything. The Telegram, Discord and Slack deployment belongs to OpenClaw, the open source runtime behind the Foundry card, and is configured per channel in OpenClaw.

SourceAIsa, aisa.one/docs llms-full.txt (Architecture, Use AIsa in OpenClaw)read 2026-09-25

Every call names its own model. AIsa's model gateway at api.aisa.one/v1 is OpenAI compatible and exposes 110+ models through the model catalog, from OpenAI, Anthropic, Gemini, DeepSeek, xAI, Moonshot and others, with native routes for Claude and Gemini. The caller picks the model per request, and the OpenClaw guide shows fallback chains and a different model per channel.

SourceAIsa, aisa.one/docs llms-full.txt (OpenAI Chat, Model catalog, Use AIsa in OpenClaw) and the homepageread 2026-09-25

Outside callers drive the platform through a documented API. AIsa publishes an OpenAPI 3.1 specification for its model and data endpoints (api.aisa.one/v1 and /apis/v1) with Bearer key auth, works with the OpenAI SDKs by swapping the base URL, ships a CLI, and publishes an Agent Card naming Bearer and x402 auth. Its MCP server (mcp.aisa.one, OAuth with dynamic client registration) is listed in the official MCP Registry.

SourceAIsa, aisa.one/llms.txt (Machine-Readable Resources, MCP), aisa.one/.well-known/agent-card.json and aisa.one/docs llms-full.txt (Authentication)read 2026-09-25

Nothing evaluates an agent. The Model Playground lets a person try a model with adjustable parameters and inspect payloads, and the model comparison page sets specs and prices side by side. That is manual trial and a comparison of specs and prices, not quality scoring of the customer's agent, and no harness, scored test cases or quality gate is documented.

SourceAIsa, aisa.one/docs llms-full.txt (Model Playground, model comparison)read 2026-09-25

AIsa has no browser or computer control of its own. The catalog resells scraping and crawling from third parties such as Firecrawl, Tavily and Oxylabs, which is another vendor's capability reached through AIsa.

SourceAIsa, aisa.one/docs llms-full.txt (Firecrawl, Tavily and Oxylabs API pages)read 2026-09-25