Metix AI

Interviews are booked onto the customer's calendar and Deel is a named partner, but no other connected system is named. Deel is the global strategic partner, framed as "one AI recruiter, hiring across 130+ countries", connecting hiring to employer of record payroll and compliance. Interviews are booked onto the customer's own calendar, since "qualified, interested candidates show up on your calendar", an authenticated write into a system Metix does not own, though the calendar provider is not named. Outreach goes to candidates through the customer's approved messages.

One named system is not an integration estate. No connector directory, ATS list or HRIS appears, and no count, and "custom integrations" is sold as a line item on the $299 Scale plan rather than described, so integration beyond the default is bespoke work. The pre-rebrand developer site claimed a single integration for LinkedIn, Indeed and major ATS platforms; that host now redirects to platform.metix.ai, a data API for people, jobs and companies with no ATS, LinkedIn or Indeed claim, so the earlier claim is unconfirmed rather than disproved.

Sourcemetix.ai/pricing, homepage and platform.metix.airead 2026-09-12

Mira, the agent, runs hiring from brief to booked interview in five named stages, with the customer setting the brief and approving each message.

Mira "turns open roles into booked interviews with AI role briefing, talent search, personalized outreach, qualification, and scheduling", and the vendor states the span as "Mira runs search through booking".

The funnel is published with its own numbers, showing the stages are real: 2,847 screened, outreach sent to 312 at 11 percent, 48 replying with interest at 15 percent, and first interviews booked across the week, each stage narrowing on a stated criterion.

A customer input and a human join point sit inside it: the customer writes the brief in a five minute setup naming must-haves and what can flex, Mira confirms it, and every outreach message is approved before it sends, so the flow pauses at a named point and resumes rather than running blind.

The participants are distinct: the hiring manager who writes the brief and approves, Mira, which plans and executes, and the candidate who is contacted and qualified, with the interview landing on a calendar.

No orchestration runtime or handoff between agents is described, and there is no builder for authoring a different flow; the sequence is the vendor's recruiting process, with the brief as its parameter.

Sourcemetix.ai homepage and /pricingread 2026-09-12

A maintained index of 900 million profiles, 90 million job postings and 30 million companies is searched through a documented query language. The corpus is explicitly maintained: "900 million profiles, 90 million job postings with 10 million live on any given day, and 30 million companies", with the jobs index described as daily active postings. That is a standing index refreshed on a cycle rather than context assembled per request, and it scales far past any context window.

The retrieval structure is documented to the field level. A Query Spec tree has named fields, match, equality and range operators, and all, any and not composers; a natural language route translates a sentence into the same tree; and a live contract route is what agents read "at run time to get field names and operators", so the schema is queryable rather than guessed. All three datasets take the same query tree.

Grounding reaches the hiring product too: Metix captures the role's real requirements upfront into a brief Mira confirms, then searches and grades candidates against that brief rather than keywords. The corpus is the vendor's own: no ATS sync or document store on the customer's side is documented on the current site, so the customer's own hiring history is not an input.

Sourceplatform.metix.ai and metix.ai homepageread 2026-09-12

Every outreach message needs the customer's approval before it is sent, and Metix only charges for candidates who clear the customer's bar. The plans FAQ states: "You write the brief and approve every outreach message before it sends. Mira then runs search through booking." It repeats it in the agency answer: "You approve outreach.

Mira sources, screens for interest, and books." Nothing reaches a candidate until a person has approved the specific message, and the vendor uses that fact to distinguish its self-serve product from its own managed service. Recruiters also review a validated shortlist before moving to conversations.

A second gate sits after the agent and is enforced economically: candidates "reach you only after clearing your bar for fit and real interest", the customer sets that bar in the brief, and if nobody clears it no credit is spent. The vendor draws the managed-service line itself: Delivery service, where "someone on our side works the role from day one", is a separate quoted purchase and not one of the four plans, and on the plans "a Metix person is not a last confirmation on the shortlist". The approval gate therefore belongs to the platform, not to an engagement.

Sourcemetix.ai/pricing FAQ and homepageread 2026-09-12

SOC 2 Type II and ISO 27001 are both shown as in progress, so no security certification is held yet, and no customer access controls are documented. The footer of every page renders a SOC 2 Type II badge and an ISO 27001 badge, each captioned "In progress" and linking to section 7 of the privacy policy; a certification in progress is not an attestation. A GDPR badge carries no such caption but is a self asserted regulatory posture rather than a third party attestation.

There are no access controls for the customer either: no single sign on or SCIM, no role based access or permission model, and no admin console appears on the hiring site or the developer platform. The nearest thing is an API key with scopes and quota shown on a status route, which authenticates a developer rather than governing who in the customer's organization may act.

Also published are Terms of Service, a Privacy Policy with a security section and a Master Service Agreement, plus a data minimization design on the data platform, where a search returns encrypted identifiers and "nothing leaves the index until you name it"; that governs what the API discloses, not who may act. Either certificate, once issued, changes the picture.

Sourcemetix.ai footer, /pricing and platform.metix.airead 2026-09-12

The agent's funnel is reported in numbers and each charged candidate is accounted for, but no record of what the agent did to whom is documented. The published view shows 2,847 screened with 11 percent reached out, outreach sent to 312 with 15 percent replying with interest, and 48 interested candidates, alongside booked interview slots by day.

Reports are a named destination in the site navigation, and the developer platform exposes a usage view and a free key status route reporting scopes and quota. On the credit model each unlocked candidate is individually accounted for, so a customer can reconcile what the agent delivered against what was charged.

Nothing keeps a record of conduct that can be reconstructed: no page names a run log, action history or audit trail, nor any export or retention period. Nothing states that a record is kept of which candidates Mira screened out and why, which messages it sent and when, or how a qualification verdict was reached. Counters answer how many; an audit trail answers what happened to whom and why. On a product whose agent contacts named individuals under the customer's brief, and which holds no security attestation, the second is the one a buyer would want.

Sourcemetix.ai homepage, /pricing and platform.metix.airead 2026-09-12

The data platform is stateless and each role starts from a fresh brief, so no agent memory is documented. The documented integration is three calls (get a key, search, read the records), with a search returning encrypted identifiers that are then exchanged for records.

Nothing in the published contract carries state between requests, and while the vendor documents rate limits, credit rules and refusal behavior in detail, it names no session, context or memory concept.

In the hiring product the brief is a per role input the customer writes in five minutes, not a store the agent builds across roles, and nothing states that Mira carries what it learned on one search into the next or that a rejection on one role informs another.

The 900 million profile index persists, but it is the vendor's data product, and credits, allowances and usage history are account bookkeeping. No named memory layer or retained preference set appears, nor any decision trace across runs, stated scope or lifetime, or customer control to inspect or purge agent state. The silence is meaningful on a documentation surface this precise about everything else.

Sourceplatform.metix.ai and metix.ai homepageread 2026-09-12

No region, data location or hosting option is named anywhere, even on pages that itemize everything else.

The hiring pricing page lists four plans down to candidate and role allowances, billing cadence and the 10 percent yearly discount, with no region, residency or hosting row at any tier, not even on the $299 Scale plan where custom integrations and an SLA are the named differentiators.

The developer platform publishes rate limits, page sizes and a changelog among much else, and names no region either. There is no region list, customer environment or way to choose, and not even a one sentence statement of where data is kept.

Self hosting, on premises and private deployment controlled by the customer are not documented. The only geography on the site is market coverage, hiring across 130+ countries through the Deel partnership, which is where candidates are rather than where data sits, and the Delaware registered office says nothing about infrastructure. Section 7 of the privacy policy, which the compliance badges link to, is the likeliest place for a residency or subprocessor commitment.

Sourcemetix.ai/pricing, homepage and platform.metix.ai; metix.airead 2026-09-12

Four installable agent skills teach an assistant to use the API, but the hiring work is done by one agent, Mira. The four skills install with a single command, npx skills add MetixAI-Official/metix-skills, with their own documentation page and an "All four skills" link that works as the catalog.

Each is a discrete unit that installs independently; the vendor states "they work in any agent, with MCP or without it", and they teach an assistant the endpoints and have it read field names and operators from a live contract route at run time. Install one and it does work; remove one and the others are unaffected. There is no browsable marketplace.

The units are skills for using the API, not prebuilt working agents. They teach a customer's assistant how to call Metix and do not themselves source, screen or write outreach. On the hiring side the product ships one agent, Mira, and the plans differ by candidate and role allowance rather than by which agents a buyer adopts. Prebuilt agents for parsing, matching and JD generation were described on the pre-rebrand platform and no longer appear in that form.

Sourceplatform.metix.ai and its skills documentation indexread 2026-09-12

Work starts when a role is defined, and candidate replies drive what happens next, but the outreach channels are named only by implication. Outreach is sent to candidates, personalized per person, and every message is approved before it sends; interviews land on the customer's own calendar, since "qualified, interested candidates show up on your calendar", with booked slots shown by day.

The customer starts from a job description file or a similar profile, or from a LinkedIn or job URL. Events drive the return leg: replies are read and classified for real interest, and a candidate clearing the bar triggers both the booking and the credit. The trigger for the whole run is the role being defined, and the vendor's claim is that a brief produces candidates ready to interview within 24 hours.

The channel set is not listed anywhere on the current site: email is evident from context rather than named, no second outbound channel is stated, and no schedule control, cadence setting or event framework reaching outside the product is documented. Continuous automated outreach and follow-ups across email and WhatsApp were described on the pre-rebrand OpenJobs site; WhatsApp no longer appears on metix.ai and is unconfirmed rather than retired.

Sourcemetix.ai homepage and /pricingread 2026-09-12

No model or provider is named anywhere, and there is no choice of model. The hiring pages, the pricing page and the developer platform show no routing description, model picker or admin setting, and no way to bring your own key, and no LLM is named for capturing the brief, matching and grading, or writing outreach and qualifying replies. The product uses its own AI for search and matching.

The developer platform documents everything else, from endpoints and query language to credit arithmetic to the quarter cent, and even notes that contact enrichment "is announced and not open, not callable, not billable"; a surface that precise, naming no model, is not withholding a choice.

Metix publishes an MCP server and four agent skills so a customer can drive it from Claude, Codex or any assistant, but that is the customer's assistant calling Metix. Choosing which assistant issues the call is not choosing the model that runs Metix's matching or writes its outreach.

Sourceplatform.metix.ai, metix.ai homepage and /pricingread 2026-09-12

A self serve developer platform publishes a documented REST API, an MCP server and four installable agent skills, all under one key. The REST API has a published Query Spec query language, natural language search and routes that fetch records by id across people, jobs and companies, backed by a quickstart and full API reference, plus an error reference and changelog.

The MCP server runs over streamable HTTP at mira-api.metix.ai/mcp, MCP 2025-03-26, with a legacy SSE endpoint alongside, and the vendor states that "the MCP surface is the REST surface and nothing more, under the same key, the same limits and the same credit rules." Four agent skills install with npx skills add MetixAI-Official/metix-skills; "they work in any agent, with MCP or without it", and read field names and operators from a live contract route at run time.

The operating contract is published rather than discovered: rate limits, page sizes and batch ceilings, the rule that a record takes two calls, and the one case where paging can miss a record are all written down. Every response arrives in the same envelope, and a query refused for an unknown field is not charged. It is self serve and ungated: a key is created in about a minute on a free plan with no card and no sales call, and the key status route is free to call. An llms.txt and a page for AI assistants are also published.

Sourceplatform.metix.ai and its documentation indexread 2026-09-12

A quality bar is applied to the agent's output and backed by pricing, but no way to test a change before it runs is documented. The homepage says candidates "reach you only after clearing your bar for fit and real interest", that "a built-in quality bar decides what reaches you", and that "quality is checked at every step" through the funnel.

The gate is enforced economically: "A credit only unlocks a candidate who's qualified and genuinely interested. If no one clears the bar, nothing is charged." A vendor that declines revenue when its agent's output fails the bar has a gate with consequences, the verdict is binary and readable per candidate, and the customer sets the threshold in the brief.

Nothing puts a change under test: no test environment or sandbox, no holdout or regression check, and no versioning or evaluator model is documented, and nothing states that a changed search, brief or message template is run against a known set before it reaches live candidates. The published figures, 95 percent scheduling within 24 hours, 5x the response rate of cold outreach and 65 percent less manual work, are outcome statistics across roles, not a comparable result for each change.

Sourcemetix.ai/pricing FAQ and homepageread 2026-09-12

Candidates are found and contacted through searches on the server and documented calls, with no browser or computer use. Search routes run against Metix's own index, records are read by id through detail routes, and outreach and booking run through the customer's approved messages and calendar. Three documented calls take a role from nothing to records, and nothing in that path opens a page. No hosted or local browser, desktop session or computer control appears on the hiring site or the developer platform, and no RPA or browser extension either.

How often Metix breaks when a target site changes its markup is a meaningless question for this product. The one thing that touches an outside interface is outreach delivery, and the site names no interface for it. A customer can paste a LinkedIn URL or a job URL when defining a role, but the person decides what is read and the product parses it, which is ingestion rather than an agent navigating.

Sourceplatform.metix.ai and metix.ai homepageread 2026-09-12