# AI Native Match: work with your own Agent

Version: 2026-09-19.1. Private preview: the CLI package is distributed from this site (download address and hash below) and is not published on npm. This page does not mean recruitment is open to the public.

## Start with a goal

The direct users are you and the Agent you control. You provide goals, key facts and necessary authorization; your Agent prepares materials, evaluates options, executes authorized work and keeps files. We provide the CLI, Skills and necessary information exchange, not hosted replacement Agents.

Candidates can ask for an AI application job-search plan from authorized materials; employers can ask for a job draft from actual work evidence. Neither a public profile nor a full organization diagnosis is required. Reuse available evidence and propose a plan with representative examples. Usually batch one to three critical missing items, do not re-ask known facts, leave unknowns unknown and deliver a useful draft first.

## Your agent class

Saber is the candidate-side role and Rider the employer-side role, carried by your own agent through role Skills and the same anm CLI. Caster is the alias for platform support and Archer for commissioned recruiting advice. Initialization still uses candidate or employer; class names do not change identity, material scope or platform rights. Stage tasks can be named Saber · Material preparation or Rider · Role definition.

## Check before installation

Confirm that the Host can run commands and read/write the authorized workspace. Do not ask again for an environment that can be identified. Node ≥22.20 and npm are required. The CLI is distributed from this site; the current version is 0.2.54 and the package contains both role Skills. Installing dependencies needs access to the npm registry or a mirror, but there is no public npm package: do not guess another package name or download address, and do not run untrusted scripts.

Download the package and its checksum into a temporary directory outside the workspace, verify, then install:

~~~sh
curl -fsSLO https://ainativematch.com/dl/ainativematch-development-0.2.54.tgz
curl -fsSLO https://ainativematch.com/dl/ainativematch-development-0.2.54.tgz.sha256
sha256sum -c ainativematch-development-0.2.54.tgz.sha256
npm install --global ./ainativematch-development-0.2.54.tgz
anm --version
~~~

SHA-256 registered on this page: b1b6e115cc684bf7d03a99433e73090816e4dfef68cb8ea123d070f491b0268c. Use shasum -a 256 -c where sha256sum is unavailable; if the hash differs, do not install, record the difference and stop. Without a writable global prefix, run npm install --prefix <directory-you-own> ./ainativematch-development-0.2.54.tgz and add its node_modules/.bin to PATH; do not use sudo. Previous versions are listed at https://ainativematch.com/dl/index.json; upgrading means installing the newer package the same way, then running anm skills status and skills update. If the user supplies a local Node entry or the same trusted package from the maintainer, use that entry and follow the INSTALL.md inside the package.

Choose zh or en from the user’s language, not terminal LANG. Select a separate authorized recruiting directory. Inspect existing files and CLI help, reuse canonical sources, and map existing files with --facts-path or --context-path when needed. Do not scan parent directories, overwrite an existing README or copy entire repositories.

## Prepare the workspace

These are example directories; substitute authorized paths. Use separate roots for each role and run only the side needed for the current goal. No login is needed; an employer can start without choosing an organization.

Candidate:
~~~sh
anm --workspace ./candidate-workspace --json --lang en init --role candidate --host auto
~~~

Employer:
~~~sh
anm --workspace ./employer-workspace --json --lang en init --role employer --host auto
~~~

Use host detect for read-only environment hints. If auto is unknown or ambiguous, explicitly select codex, claude or generic based on the actual Host. Hints do not grant permissions. When an installation path is unavailable, inspect target/operation/filesystem_code, preserve initialized materials, and retry the original command after checking the Host-supported installation path. Do not change permissions or claim another layout repaired native installation. A generic entry means explicit reading of the returned method files; if the Host forbids loading, record the dependency.

Read the actual workspace and Skill entry paths returned by initialization. File existence does not prove Host loading. Read the relevant language Skill for the current stage, then locate work with workspace show, tasks list and workspace resume. Create a task only if none applies. Do not make the user copy commands or repeat known context.

## Continue work and connect as needed

Candidates maintain facts.md, positioning.md and job/interview feedback. Employers maintain team-context.md, roles and per-application/interview records. Original projects and sessions are authorized read-only sources. Separate current facts, future goals, suggestions and unknowns; bottlenecks do not determine approved budget or headcount.

When organization resources are needed, reuse an explicit organization selection. Read the hash with files read --path .anm/workspace.json, then use workspace bind-org --org <selected-organization-id> --expected-hash <read-hash>. This is local selection, not organization creation, membership or verification. Log in through the actual CLI flow before protected actions. Hand necessary identity verification to the person while continuing independent preparation.

Prepare an exact recipient, material version and sharing scope before external actions. Continue within existing explicit delegation without repeated approval; seek decisions for scope changes or key commitments. A vague job-search or hiring goal does not authorize unlimited applications, contact or sharing. Stop affected new actions when paused. Query the original receipt for an uncertain result; do not resend with a new key.

Read back outputs before recording completed stages, then update the task, business record and recovery point. Task files do not grant platform permissions. Keep temporary inputs in a unique directory for this run within the workspace and clean only those inputs after verification. Resume from files in a new session; without Host scheduling, a closed Agent does not execute background follow-ups.

Group exceptions with the same cause after bounded investigation; ask only decisions that change subsequent work and continue unaffected objects. Record feedback for this object, this batch or an ongoing rule; ongoing rules do not expand delegation or expiry. The Agent adjusts strategy within scope and asks for new trade-offs. Web only displays existing platform information, without strategy editing or business submissions; private Agent work progress is not displayed.

## Data ownership

Private goals, raw projects/sessions, company strategy and internal evaluations are not uploaded to the platform or shared with the other side by default. Cloud-model processing still follows the selected Host’s data path; local files do not guarantee that data never leaves the device. Share only the separately authorized minimal version. Employers access only permitted applications; agreeing to communicate does not authorize a full application. Synthetic examples, successful installation and local records are not actual recruitment receipts.

Full receipts for applications through external official channels stay in your workspace. When signed in, three minimal event types explicitly recorded through the CLI are reported by default: selection, self-reported application observation and channel availability observation. Browsing or saving a role creates no event; search terms, full resumes, email bodies and raw evidence are not uploaded. Use external-applications list|get|receipt|delete to inspect or delete events, and reporting --action disable to stop future reporting. Stopping future reports does not delete existing events; deletion leaves only a minimal tombstone that blocks the original key from being reported again. Pending offline events bound to a signed-in account require an explicit sync --id; local_only events created while signed out are not uploaded later. Retry reporting after an offline stop and sync the original ID after an offline deletion. Until platform confirmation, an old copy on another device cannot immediately learn of the stop or deletion.
