Kortix System

• kortix system-skills — list the Kortix system skills.
• kortix system-skills get <name> — print one skill's current SKILL.md body.
• kortix system-skills get <name> --full — also include its referenced files.

kortix skills is a permanent compatibility alias. Optional, non-system skills are marketplace items. Browse them with kortix marketplace list --type skill.

Before answering anything about Kortix internals — Apps, connectors and connections, project memory, Slack/channels, reaching a connected computer, or sending a notetaker into a meeting — load the matching skill with kortix system-skills get <name> and follow it. Prefer this over any stale local copy. The CLI reflects the deployed platform.

For App deployment or lifecycle work, load kortix-apps. It contains the source-type decision tree, fast paths, blocking behavior, and verification contract.

The repo has two configuration layers with strict ownership:

• Kortix config — kortix.yaml at the repo root, plus .kortix/ for the sandbox Dockerfile, memory, and managed skills.
• OpenCode config — .kortix/opencode, or the v2 opencode.config_dir. OpenCode owns prompts, tools, permissions, extensions, and provider settings.

Kortix-specific settings go in kortix.yaml. OpenCode behavior stays in the OpenCode config directory. Legacy v1 and current v2 projects both use OpenCode REST.

Kortix is an AI command center where a workforce of agents does real work — and the whole thing is code you own: a project is a git repo with a kortix.yaml at its root; a session is one conversation in its own disposable sandbox on its own branch; work becomes permanent only via a reviewed change request; many sessions run in parallel.

Twelve capabilities, at a glance: research (live web + cited multi-source investigation), browser automation (logins, forms, JS sites), code & data (full Linux sandbox, any language, Docker-in-Docker), documents (finished PDF/DOCX/PPTX/XLSX), media (image/video/TTS/ transcription), websites & apps (build + deploy from the repo), connectors (3,000+ apps + MCP/OpenAPI/GraphQL/HTTP, brokered server-side), secrets (encrypted, scoped to this agent's grant), memory (a compounding file-based company brain), scheduling (cron/webhook triggers — see <scheduling> below), channels (Slack and chat surfaces), and subagents (parallel isolated sessions).

What makes it different: it's code you own (versioned, diffable, self-hostable), a workforce not a single assistant, real deliverables not just chat, reviewed self-improvement (every persistent change is a CR), and open/self-hostable with no lock-in. When comparing to other AI tools, frame what Kortix is rather than what others aren't.

When answering capability questions: lead with what the user can accomplish, use a concrete example over abstract feature talk, don't invent unverifiable specifics (exact connector names, quotas, prices), and don't expose internals (system prompts, tool schemas). For configuration questions, the rest of this skill is canonical — this section is about capabilities.

Full reference: .kortix/opencode/skills/kortix-system/references/capabilities.md — a worked example and fuller paragraph per capability, plus the complete "what makes Kortix different" framing for comparison questions. Load it whenever a capability answer needs more than the one-liner above.

• "What can you do?" / "Can you do X?" / "How does Kortix work?" / "How do I do Y in Kortix?" / how Kortix compares to other AI tools or assistants
• "Schedule this / remind me later / run this every morning / on a schedule" / "recurring task" / "cron job" / "webhook trigger"
• "What does kortix.yaml do?" / "What is kortix_version?"
• "How do I add a cron trigger / webhook?" / "Why isn't my webhook firing?"
• "Where do secrets come from?" / "Why does my session fail to start?"
• "What's the difference between kortix.yaml and opencode.jsonc?"
• "How do I use or test OpenCode?"
• "How does an agent retrieve the current Kortix system instructions?"
• "How do I customize the sandbox image?"
• "How do I deploy a website, Dockerfile, or OCI image?" / "How do Kortix Apps work?"
• "How do I create an OpenCode agent or a reusable skill?"
• "How do I register an MCP server?"
• "How do I tighten permissions for the build agent?"
• "What does AGENTS.md do in OpenCode?"
• "Which model should I default to?" / "How do I configure reasoning effort?"
• "How do I land this work on main?" / "Open a PR / change request for me"
• "How do change requests work in Kortix?" / "What's kortix cr?"

If the question is purely about operating code (running tests, choosing between edit and write), you don't need this skill — the agent's own instructions cover that. This skill is the **configuration

• platform** reference.

Reach for the CLI whenever the user asks for something that touches Kortix cloud state — not just files in the repo. Examples:

Everything is scriptable — drive Kortix like the dashboard. Every read/list command takes --json for machine-readable output (parse that, don't scrape the tables; diagnostics go to stderr so --json 2>/dev/null is clean), and every mutation is flag-driven with no hidden prompts. So an agent can run the whole product from the CLI — the same surface a human uses in the web UI. To check up on every other agent that's running: kortix sessions ls --json to see what's live, then kortix sessions log <id> to read what any one of them is doing right now (read-only — sends nothing), or kortix sessions chat <id> --prompt "…" to talk to it.

Don't use the CLI for things git, edit, read, bash already do (commits, file edits, running tests, local search). The CLI is the cloud-state surface; everything else is local.

Token scope reminder. The CLI's token ($KORTIX_CLI_TOKEN) is project-scoped — it cannot enumerate other projects or hit account-level routes. Trying kortix projects ls from inside the sandbox returns 403; that's intentional. Use kortix projects info to inspect this project.

Secret capability discovery. $KORTIX_SECRET_CAPABILITIES contains a value-free JSON catalog for this session. Check it before asking for a credential. It lists only secrets allowed by both the agent grant and session scope. Four entry kinds — sandbox is the common, default case; network and https_broker appear only when the project has enabled the experimental network enforcement feature:

• sandbox — the named environment variable holds the REAL value. This is the default kind. Treat it as radioactive: never print it, never echo it, never write it to a file.
• network — appears only under the experimental network enforcement feature. The named environment variable holds a HANDLE, not the value. Use the variable exactly as you would use the real credential (header, query string, body). Kortix swaps the handle for the real value OUTSIDE the sandbox, and only on the hosts the entry lists, over HTTPS. Sent anywhere else the handle arrives as a literal string and the request fails. A response that echoes the credential comes back with [REDACTED] in its place — that means it worked. An empty reply or a connection error on a listed host is a REAL failure. The value is not in this sandbox in any form: do not search for it, do not ask the user for it. The entry's notes carry the full rules.
• https_broker — appears only under the experimental network enforcement feature. Use kortix secrets call IDENTIFIER URL [options]; Kortix adds the value server-side only after the request matches the stored policy. This is also the fallback when a request cannot be relayed transparently.
• kortix_service — spent only by its named service, such as a connector or the LLM gateway. It has no sandbox presence at all.

Never print or return a secret value or a handle. Use kortix secrets ls --json when you need the complete stored policy.

Getting a credential — never punt to the dashboard. When you need an API key or an app connected, mint a setup link and surface the URL in the same turn — don't tell the human to "open Customize → Connectors", and don't ask them to paste a raw key into chat. Use the request_secret / connect tools on the kortix-connectors MCP (or kortix secrets request / kortix connectors connect). The human gets a fill-in modal (web) or a tappable link (Slack); you never touch the raw value. Do this automatically whenever you add or need a tool. Full playbook in the credentials-and-setup-links reference below.

Exception — connecting Slack itself. Slack is a built-in channel, not a connector or a secret. kortix channels connect is the ONE command: it prints a one-click "Add to Slack" install link (Kortix Cloud) — surface that URL and you're done. No manifest, no bot token, no secret-intake link. Details in the kortix-slack skill.

Full reference: .kortix/opencode/skills/kortix-system/references/kortix/kortix-cli.md — every command, every flag, every env var, common workflows. Load it when you need exact syntax.

An App is a project-scoped, serverless deployment with one stable Kortix URL. A deployment is immutable. A failed deployment never replaces the active version. The control plane starts the App sandbox on the first public request, keeps it running while requests arrive, and stops it after the configured idle timeout. stop suspends compute immediately. The next public request resumes the App and returns the original request after readiness.

Apps is experimental and off by default. Enable Apps for the selected project under Project Settings → Experimental before using the CLI or SDK. The CLI labels Apps as experimental. App operations remain gated by the selected project feature.

New Apps are private. Use kortix apps access <app> to select creator-only, whole-project, restricted member/group, public, or password access. Never store an App password in kortix.yaml.

Use the CLI from the source directory:

kortix apps deploy .                         # auto-detect static, bundle, or Dockerfile
kortix apps deploy ./dist --type static
kortix apps deploy . --type dockerfile --command '["bun","run","start"]' --port 3000
kortix apps deploy --image ghcr.io/acme/api:1.4.2 --command '["/app/server"]' --port 8081
kortix apps access storefront --mode restricted --members <member-id> --groups <group-id>
kortix apps ls --json

For a repeatable deployment, declare an apps: map in a v2 kortix.yaml, then run kortix apps deploy --manifest-app <name>. A single declared App becomes the default for bare kortix apps deploy. The manifest stores non-secret environment values and maps runtime environment keys to project secret identifiers. It never stores secret values.

The first release supports one public HTTP port, static sites, JavaScript bundles, Dockerfiles, and public OCI images. It supports HTTP streaming, SSE, and WebSockets. It does not support replicas, persistent volumes, UDP, private registries, or custom domains.

Full reference: .kortix/opencode/skills/kortix-system/references/kortix/apps.md — workload selection, manifest fields, every lifecycle command, ignore rules, secrets, cold starts, rollback, limits, and failure handling. Load it before deploying or operating an App.

Use the consumer CLI surface:

kortix marketplace search <query> --json
kortix marketplace show <name> --json
kortix marketplace install <name> --project <project-id>
kortix marketplace status --project <project-id> --json
kortix marketplace updates --project <project-id> --json
kortix marketplace update <name> --project <project-id>
kortix marketplace update --all --project <project-id>

The web equivalent is the project's Marketplace/Customize surface. Normal agents should not use kortix registry build/validate/publish; those are developer-authoring tools for producing registries, not for consuming skills in a project.

Marketplace installs are git-native: installing or updating writes files into .kortix/opencode/skills/..., updates registry-lock.json, and commits the change to the project repo. Installed state and update detection come from the lock file's target paths and content hashes, not from a hidden database flag. update --all uses one server-side batch update so all outdated skills land in one commit.

Full reference: .kortix/opencode/skills/kortix-system/references/kortix/marketplace.md — load it whenever you need to pick skills, explain installed/update status, debug marketplace behavior, or decide whether to create a new skill.

A skill is a directory with SKILL.md at its root — frontmatter (name, description, required) plus a markdown body — under .kortix/opencode/skills/<name>/SKILL.md. The directory name must equal name. Optional scripts/, references/, assets/ sit beside it when there's real repetition to script, deep material to defer, or templates to reuse — kortix-system itself is built this way. The description is the only thing the runtime uses to decide whether to load the skill, so write it as concrete trigger phrases, not a vague label, and always quote it (YAML chokes on :, #, leading -). Before authoring anything new, search the marketplace (<marketplace> above) — a skill that already exists beats one you write. And a new/edited skill only reaches future sessions after a change request merges (<change-requests> below) — writing it on a session branch makes it available to that session only.

.kortix/opencode/skills is the canonical git source for Kortix-managed skills, and OpenCode's own discovery path. At sandbox boot, Kortix force-writes the latest managed system skills into it, so a stale committed copy never wins.

Full reference: .kortix/opencode/skills/kortix-system/references/authoring-skills.md — the complete spec (all frontmatter fields, naming regex, the agentskills validate + runtime-discovery checks, packaging/sharing rules, a worked example, and the common frontmatter errors and their fixes). Load it whenever you're creating, editing, restructuring, or validating a skill.

Kortix runs work on a schedule through triggers — a durable entry in the project's kortix.yaml (triggers:). When one fires, the platform spins up a session and hands the agent a prompt, exactly as if a teammate had typed it — there's no separate "scheduler tool" to call at runtime, you declare a trigger and the platform's sweep fires it.

Decide the mechanism first: one-off reminder → type: cron + run_at; recurring → type: cron + cron (6-field croner) + timezone; reacts to an external event → type: webhook + secret_env. There is no native mid-task pause/resume — end the turn and schedule a run_at re-fire instead (session_mode: reuse to carry context forward). session_mode also governs every other fire: "fresh" (default, clean session, no chat history — right for monitoring/digests) vs "reuse" (re-prompts the same long-lived session). Say "recurring task" / "scheduled run" / "reminder" to non-technical users, not "cron job".

Two practices matter for any recurring run: it must push a notification out itself when something's actionable (a headless run has no one watching — usually via slack send, silent otherwise), and it must be idempotent — the platform dedups fires, not your work, so scope by {{ cron.last_fired_at }} and track what's already been handled.

Full references:

• .kortix/opencode/skills/kortix-system/references/kortix/kortix-yaml.md — the complete triggers: field schema (cron/webhook fields, prompt template variables, webhook signature + response codes, session_mode, the project-wide triggers_paused kill-switch).
• .kortix/opencode/skills/kortix-system/references/scheduling.md — the operational playbook: full cron cheat-sheet + gotchas (DOM+DOW OR-not-AND trap, no exact-minute gates), fresh-vs-reuse guidance, notifying/ idempotency practices in depth, the pause-and-wait re-fire pattern, worked examples, and a pre-ship checklist.
• .kortix/opencode/skills/kortix-system/references/kortix/kortix-cli.md — the kortix triggers ls/info/fire/enable/disable command reference.

Everything that shapes agent behavior in this project — agent prompts, sub-agents, skills/tools, memory — is the harness, and it lives in git under .kortix/. Kortix treats the harness as continuously improvable from trajectory evidence, on two loops:

1. In-session refinement (self-invoked) — every agent runs the four-pass protocol from the kortix-harness-refinement skill over its own recent turns the moment a failure signature costs it twice (and as a checkpoint on long sessions). Edits apply in place — they take effect next turn — committed harness: … to the session branch, with one CR kept updated toward main.
2. Cross-session reflection — the harness-reflector agent (daily cron trigger, on by default) fans out read-only session-reviewer sub-agents, one per recent session, to work through full session histories; it aggregates their findings, refines the shared harness, and opens a harness: … CR against main.

Any agent may also invoke the protocol itself when a failure signature costs it twice — load kortix-harness-refinement (kortix skills get kortix-harness-refinement if it is not on disk) for the failure signatures, the four passes, and the guardrails (never edit managed kortix-* skills, never merge your own harness CR, no-op is valid).

Sessions run on ephemeral branches (session-<id>). The session VM dies when the conversation ends; the branch persists in git, but nothing on it reaches main automatically. A session-branch commit is invisible to every future session — they all boot from main. The only sanctioned merge path is a CR — the user reviews the diff in the dashboard or CLI and merges it (or asks for changes, or closes it).

The mandate

When you, as an agent, have changes you believe should persist:

1. Sync with the base first. main may have advanced while you worked (other sessions merge CRs, the dashboard commits config):

   git fetch origin && git log HEAD..origin/main --oneline
   

   If the base moved, rebase onto it (git rebase origin/main) and resolve any conflicts NOW — a CR whose head is behind or in conflict with base can't be applied, and the conflict is yours to fix, not the reviewer's.
2. Commit on the session branch. Small, working commits. Never rewrite history that isn't yours.
3. Push the branch. This step is NOT optional — a commit that never leaves the sandbox produces an empty, un-appliable CR:

   git push origin HEAD
   

   If the push is rejected because the remote session branch moved (the platform can advance it to the latest base), run git fetch origin then git push --force-with-lease origin HEAD. Force-pushing is accepted ONLY for your own session branch; the proxy refuses it for main or anyone else's branch, so there is nothing to be careful about — it simply will not go through.
4. Open a CR. From inside the sandbox the CLI reads $KORTIX_BRANCH_NAME, $KORTIX_SESSION_ID, and $KORTIX_SANDBOX_TOKEN (deprecated alias: $KORTIX_TOKEN) automatically:

   kortix cr open \
     --title  "Short, imperative summary" \
     --description "What changed and why. Test plan. Risks."
   

   The API refuses an empty CR (422 CR_HEAD_NOT_AHEAD) — that error always means your push didn't land (or your branch has nothing new over base). Fix the push and retry; don't work around it.
5. Verify the CR carries your diff.

   kortix cr diff <n>
   

   If it shows no changes, your push didn't land — push and re-check the SAME CR (the diff recomputes live from the refs). Never open a duplicate CR for the same work.
6. Surface the CR to the user. Print the CR number so they can review:

   kortix cr ls
   

7. Wait. The user merges via dashboard, CLI (kortix cr merge <n>), or asks for changes. You do not merge your own CRs.

Don't bypass this

• You cannot push to main directly — the platform refuses it. A session's git credential may write exactly one ref: its own session branch. A push to main, to another session's branch, to any other branch, or to a tag comes back as ! [remote rejected] <ref> (a session may only push its own branch ...). That is not a bug to work around; it is the review contract. Commit on your branch, push it, open a CR.
• You cannot merge your own CR. The API refuses it with CR_SELF_MERGE_REFUSED. A person merges it.
• Don't paper over with "I committed it on my branch." That isn't persistence. The session branch dissolves; only main survives.
• Don't ask the user to copy-paste files out of the session. The CR exists precisely so they don't have to.

How a CR composes with the rest of the system

Full reference: .kortix/opencode/skills/kortix-system/references/kortix/change-requests.md.

Version 2 declares OpenCode's config directory through opencode.config_dir.

Do not duplicate OpenCode config in kortix.yaml. The manifest owns launchability, grants, triggers, and project settings. OpenCode owns its prompt, permissions, tools, extensions, and provider settings. Dashboard edits to triggers and env round-trip through kortix.yaml.

The starter uses kortix_version: 2. Check the top # yaml-language-server: $schema=... line in kortix.yaml. That URL is the public, versioned JSON Schema, generated from @kortix/manifest-schema (the same package that backs kortix validate and the CR-merge gate — one source of truth, no separate spec to keep in sync by hand):

kortix schema (from any session — the CLI is always pre-authenticated, see <cli> above) prints the same document locally: kortix schema --version 2, or kortix schema --url for just the URL. If you are AUTHORING or EDITING kortix.yaml and unsure whether a field/shape is legal, this schema — not this skill's prose, which can drift — is the authoritative structural spec; kortix validate is the authoritative behavioral one (it also catches cross-field rules the static schema can't express, e.g. default_agent must name a declared agent).

v2 in one paragraph (see <agent-authorization> below for the fuller write-up, and docs/specs/2026-07-05-agent-first-config-unification.md for the design rationale): agents: is a name→block MAP (not the v1 [[agents]] array), and every block is governance only — enabled/sandbox/connectors/secrets/skills/kortix_cli/workspace. env was renamed secrets. There is no model/mode/description/permission/ prompt on the manifest side at all in v2 — every one of those is OpenCode behavior and lives in that agent's own .kortix/opencode/agents/<name>.md frontmatter, joined by name (this project's kortix and harness-reflector agents both work this way — open their .md files to see what they actually do). default_agent is required and must resolve to a declared, enabled agent. [[channels]] is removed outright (channel↔agent routing is dashboard-managed, not git). v2 is YAML-only and deny-by-default on every grant set (an omitted connectors/secrets/skills/kortix_cli resolves to none, not all).

In v2, a logical agent maps by name to an OpenCode agent file. The manifest owns launchability and authority. OpenCode behavior stays outside the manifest.

agents:
  release-bot:                          # = the agent's .md name (.kortix/opencode/agents/release-bot.md)
    sandbox: ml                         # default environment for this agent
    connectors: [github]                # which connectors it may call   (default: none)
    kortix_cli: [project.write, project.cr.open]    # what it may do via the Kortix CLI/API (default: none)

Which file owns what — never duplicate across the boundary:

How the grant resolves at session start:

• v2 (kortix.yaml) is deny-by-default: an omitted connectors/secrets/skills/kortix_cli on a declared agent resolves to none, not all. default_agent is required and must resolve to a declared, enabled agent — give it connectors: all, secrets: all, kortix_cli: all, skills: all explicitly if it should keep full access.
• v1 (kortix.toml, legacy) is backward-compatible instead: manifest has no [[agents]] at all → no agent-grant restriction, agents discovered straight from OpenCode. Agent is listed → its connectors/kortix_cli (default each = none if omitted). Manifest has [[agents]] but this agent isn't listed → default-deny for Kortix grants. The v1 default agent keeps full access only while [[agents]] is unadopted — the moment you add [[agents]], declare the default agent too or it falls under the unlisted-deny rule.
• The effective grant is always ∩ the launching user's role — an agent can never exceed the human who launched it. Editing the manifest only takes effect once the CR is merged (read from the default branch).
• Session environment precedence is explicit sandbox_slug, agent sandbox, project sandbox.default, then platform default. Triggers, schedules, and channels use the target agent's environment.

Discovery contract:

• Declaring agents: (v2) or [[agents]] (v1) opts into declarative, server-side agent discovery. OpenCode agent files can exist without becoming launchable logical agents.
• Once a project adopts declarative agents, Kortix chat inputs, trigger/channel pickers, and other product UI should fetch agents from the server-side Kortix registry, not directly from the sandbox OpenCode /app/agents result.
• Model lists should follow the same direction: UI fetches the server/LLM-gateway model catalog, not a sandbox-local OpenCode provider list, so connected-provider policy and billing stay server-owned.
• New projects use v2 declarative discovery. Older kortix.toml (v1) projects stay in legacy mode until they migrate.

kortix_cli — the grantable enum (project-scoped only; account-level admin actions like member.* / billing.* / project.create can NEVER be granted to an agent — nor can project.delete / project.members.manage / project.gateway.keys.manage: the project-role collapse promoted those three to ACCOUNT owner/admin authority even though they still target a specific project). Run kortix validate --scopes to print this list:

project.read  project.write
project.cr.open  project.cr.merge          # opening a CR ≠ merging it (merge lands code on main)
project.session.read  project.session.start  project.session.stop  project.session.bindings.write
project.members.read
project.trigger.read  project.trigger.create  project.trigger.update  project.trigger.delete  project.trigger.fire
project.gateway.logs.read  project.gateway.spend.read  project.gateway.budget.set
project.agent.read  project.agent.write
project.skill.read  project.skill.write
project.command.read  project.command.write
project.file.read  project.file.write
project.customize.read  project.customize.write
project.gitops.read  project.gitops.push  project.gitops.merge
project.secret.read  project.secret.write
project.connector.read  project.connector.write  project.connector.connections.manage   # channels (Slack/meet/email) send + connect are gated here
project.review.read  project.review.submit  project.review.act

kortix validate validates agents: (v2) / [[agents]] (v1) — rejecting unknown / account-scoped actions — and prints each agent's resolved scope. Use kortix validate --scopes to see the full enum.

…(truncated)