FAQ
What Nous AI gives your AI agents, how to connect it, what the benchmark shows and how data is handled.
Updated
Nous (taicheck.com) is an intelligence layer for AI agents: it gives an agent the professional skills, tool guidance and quality checks a task needs, through one MCP connection. An agent asks Nous for the skill that fits the job in front of it, follows its steps and template, and can have Nous check the result before handing it over.
How Nous works →Nous is for developers and solo founders who run AI agents on professional work, such as a finance report, a campaign plan or a product spec. It is not a hiring, recruiting or salary product, and it never judges or ranks people.
A Nous skill is a SKILL.md for a role or an area of expertise, such as finance or user research. Each one sets out the steps for a task, the profession's own standards, the template for the deliverable, the checks it must pass, common pitfalls, and the tools the work is done with. Skills are versioned, so an agent can tell when one has changed.
Nous connects over MCP (the Model Context Protocol) at https://www.taicheck.com/api/v1/intelligence. Sign in, create an API key under Dashboard → Connect, keep it in an environment variable, and add the endpoint to your agent. In Claude Code that is one command: claude mcp add --transport http nous https://www.taicheck.com/api/v1/intelligence --header "Authorization: Bearer $NOUS_API_KEY".
Developer guide →Nous works with any agent or client that speaks MCP over HTTP, such as Claude Code, and its guidance adapts to the model and tools available. Nous's own benchmark has so far tested three small, fast models: gpt-oss 20B, Gemma 4 and Claude Haiku 4.5. Larger models have not been benchmarked yet.
The main MCP tools are tai_get_skill (the skill for a task), tai_context (what the agent needs for one step), tai_check_output (checks a result against the skill's checks and says accept, retry with fixes, or block), tai_skill_tools (how to reach the tools the work needs), tai_find_assets (openly licensed pictures with their credit lines, and open sources for icons, fonts, wireframe kits, templates and data), tai_check_updates (which skills changed) and tai_skill_feedback (tell Nous how a skill worked).
In Nous's Harbor benchmark, the same agents on the same tasks met 88% of the task checks with Nous and 74% without, a gain of 14 points (95% interval +9.4 to +18.5). That is the closed standard v2 round: 38 paired runs, run by Nous on three small, fast models (gpt-oss 20B +15, Gemma 4 +14, Claude Haiku 4.5 +13), closed 1 November 2026. Every model tested improved, and the next round (v3, on the rewritten skills) is running.
Live results and methodology →Each run gives one model the same work task twice, once with the Nous skill and once without, with the same instruction and starting files. Both deliverables are scored by the task's fixed checks, not by opinion. The tasks, checks, scoring and harness are fingerprinted and frozen for a round; any change starts a new round. Results show wins, losses and a 95% interval, and losses stay on the page.
Benchmark methodology →Nous writes its skills from public job postings, the O*NET and ESCO occupation frameworks, and public guidance from regulators and standards bodies where the licence allows commercial use. It publishes patterns, never individual records: a job-postings figure needs at least 20 postings from at least 5 employers. Professionals also rate example answers, and those ratings tune the checks.
Nous has published skills for marketing, sales, finance, design, product and leadership, plus areas of expertise such as data analysis, user research, accounting, teaching and software engineering. The evidence is uneven: the capabilities page shows, for each area, how many job postings its skill is built from and how many benchmark runs it has.
Capabilities and live status →No. When an agent sends a result to tai_check_output, Nous reads it in memory, checks it and drops it. It keeps only the outcome: the score, which checks failed, the action taken and the time, never the text itself.
No. Nous serves patterns, not records: what a profession's work involves, drawn from many sources at once. It does not sell or export raw postings, profiles or personal data, and it is never used for salary decisions or to judge people.
Nous is priced by the tasks your AI gets done each month: Starter $19 a month (About 50 tasks a month), Builder $49 a month (About 150 tasks a month) and Pro $199 a month (About 750 tasks a month), and Enterprise at custom pricing. A task is one piece of work, such as a campaign plan or a budget review, and uses about 10 credits: 5 to find the skill, 1 for each piece of guidance and 3 to check the result; checking for updates and sending feedback are free. When a month's tasks run out, Nous pauses until the next month or a larger plan, with no extra charges.
Plans →Yes. Ten rounds on Contribute within seven days, rating two answers to a task from your field or finishing short scenarios, earn a free month of the Starter plan, once every six months. You can use Nous on your own tasks during that week.