Junis kit

imageCI imageLicense: Apache-2.0 imageLatest tag

한국어

A skill and a local MCP bridge that let your coding agent build, run and fix squads of AI agents on Junis.

You describe the squad in Claude Code, Codex, Cursor, Gemini CLI, OpenCode or GitHub Copilot. The agent reads your organization first, takes model and tool names from the platform's live lists, checks each instruction before saving it, runs the squad once and reports what it changed, what it assumed and what is left for you. The squad then keeps running on Junis, on its schedule, with its own models, knowledge graph, memory, wiki and tool-call log, after your coding session ends.

Install

Requires Node.js 18 or newer, git, and a junis.ai account on a Basic or Pro plan.

Claude Code, as a plugin:

claude plugin marketplace add TOOEASY-crew/junis-kit
claude plugin install junis@junis-kit --scope user

Every other supported agent on this machine, pinned to a release tag:

curl -fsSL https://raw.githubusercontent.com/TOOEASY-crew/junis-kit/v0.1.0/install.sh | JUNIS_KIT_REF=v0.1.0 sh

The latest commit:

curl -fsSL https://raw.githubusercontent.com/TOOEASY-crew/junis-kit/main/install.sh | sh

install.sh registers the agents it finds. --agents takes a list of ids or all: Codex (codex), Cursor (cursor), Gemini CLI (gemini-cli), VS Code with Copilot (vscode), GitHub Copilot CLI (copilot-cli), OpenCode (opencode), Devin Desktop or Windsurf (devin-desktop), Claude Desktop (claude-desktop, which also needs ~/.junis-kit/dist/junis-skill.zip, built by the installer, uploaded once under Customize > Skills) and Claude Code (claude-code).

# Print the plan and change no agent configuration
curl -fsSL https://raw.githubusercontent.com/TOOEASY-crew/junis-kit/v0.1.0/install.sh | JUNIS_KIT_REF=v0.1.0 sh -s -- --dry-run
# Only some agents
curl -fsSL https://raw.githubusercontent.com/TOOEASY-crew/junis-kit/v0.1.0/install.sh | JUNIS_KIT_REF=v0.1.0 sh -s -- --agents codex,cursor
# Skill only, for any agent that reads Agent Skills
npx skills add TOOEASY-crew/junis-kit -g

Piped, the script clones the kit into ~/.junis-kit/src (or updates that clone) and runs the installer. From a checkout, sh install.sh installs the checkout. A skill-only install has no bridge and no plugin hook: the skill uses its own command line, and confirmations rely on the skill's rules and your agent's approval for shell commands.

The plugin follows main: update it with claude plugin marketplace update junis-kit, then claude plugin update junis@junis-kit. Update an install.sh install by running it again. The kit holds your Junis key, so read CHANGELOG.md before moving to a newer tag.

To uninstall, run claude plugin uninstall junis@junis-kit or node ~/.junis-kit/skills/junis/scripts/junis.mjs uninstall (--purge also deletes stored keys, --dry-run shows the plan). Then delete ~/.junis-kit to remove the kit files, and revoke the key at https://junis.ai/settings/api-keys. Deleting local files does not revoke it.

Connect

The API key is created in the first conversation. Open a new session and ask, in any language:

You:    Check my Junis connection
Agent:  Not connected yet. Open https://junis.ai/auth/api-key?code=…, sign in,
        check the permissions the kit asks for and press Approve.
You:    Approved
Agent:  Connected. The key reaches two organizations: Globex (OWNER) and
        GlobexSales (ADMIN). Which one should I work in?

Nothing is copied or pasted, so this also works over SSH and in containers.

Alternatively, create a key at https://junis.ai/settings/api-keys with the "Coding agents (recommended)" preset, press its copy button and tell the agent it is copied; the kit reads, verifies and stores it, then clears the clipboard. Never paste a key into the chat. From a terminal, run junis.mjs login --browser. In CI, set JUNIS_API_KEY.

Usage

Build a squad that summarizes https://example.com/blog every Monday, and run it once now

The agent works through the steps described at the top of this page and ends with its report.

Examples

Requests the platform supports today, written as you would type them. They are examples, not customer stories.

• "Load our supplier contracts and the parts list. Which products depend on Northwind Metals, and what do the contracts say about a supply failure?" Knowledge turns the contracts into a graph of suppliers, parts, products and clauses. The orchestrator follows it to the affected products, quotes each clause with its document and page, and answers quantity questions from the parts database through query templates your team approved.
• "Every hour, check new Sentry issues for the checkout service, test a fix in a sandbox and open a pull request. Merging stays with me." A triage agent on the Sentry server, a fixer with the GitHub server and a code sandbox, and a reviewer on a different model family in a review loop. The merge tool goes on the agent's confirmation list, which is off until you set it, and a scheduled run never calls a tool on that list.
• "Every Monday, find out what changed in EU battery regulation, update our wiki page and show me what changed since last week." Three to five parallel researchers search the web at the depth you set, a synthesizer merges their findings with sources, and a reviewer loops until the draft passes or the loop limit is reached. The page keeps its version history, so last week's text and the difference are one call away.
• "Every weekday morning, brief me from my Gmail and Linear, and remember what I decide." Gmail and Linear connect through OAuth in the web app. Memory keeps your profile, the facts you ask it to keep, a daily log and monthly, quarterly and yearly summaries; past conversations are searchable by date.
• "Every night, run the tests in ~/work/shop on my laptop and send me only the failures." npx junis links your computer, and a sub-agent runs commands, reads files and drives a browser there with your rights. With the device tool marked as required, an offline laptop marks the run as missing its tool instead of reporting an empty success.
• "Let the maintenance agent open tickets in our maintenance system, but nothing goes out until I have approved it." The agent drafts a connection and one typed action and shows the exact request it would send. You approve the draft in the web app; from then on the maintenance agent has a single tool that fills in the equipment, the summary and the priority and nothing else. A run marked for confirmation waits for a person, and every run is listed with who asked and what came back. Until your system's key is registered, the same flow runs on a simulated connection that sends nothing.
• "Answer customer questions on help.example.com from these 40 help articles." The articles go into Knowledge, marked as shared publicly. Once you switch on public embedding in the web app, the chat widget's knowledge search sees only those sources. It runs only on the domains you allow, with a message limit per visitor and without the tools you block.

Capabilities

Counts are as of November 2026. The kit reads every list live, so new models and tools work without a kit update.

Details for each area are in docs/CAPABILITIES.md.

86 built-in tools, including:

55 chat models, set per agent:

How it works

With the hosted URL alone, each agent's configuration would hold the key in plain text, the full tool catalogue would load into every session, and long runs would exceed client timeouts. The bridge keeps the key in one file, shapes the tool list per agent and returns a session id to follow. Details are in docs/ARCHITECTURE.md.

skills/junis/scripts/junis.mjs and its lib/ modules are the command line, the bridge and the installer. They use only Node.js built-in modules.

Claude Code gets full mode: every Junis tool is listed, and Claude Code loads a schema only when the tool is used. The others load every schema up front, so they get lazy mode: 11 kit tools that find platform tools on demand. --full or --lazy overrides the choice.

The installer copies the skill to ~/.junis-kit/skills/junis, links it into each agent's skills folder and adds an MCP server named junis to each agent's configuration, backing every file up first. It never rewrites a file it cannot parse, and unless you pass --force it leaves alone files with comments and any junis entry it did not write. With the Claude Code plugin installed, it skips Claude Code so tools do not appear twice.

Security

• The key is stored in ~/.junis-kit/credentials.json (mode 0600 in a 0700 folder) or read from JUNIS_API_KEY. It is never written to an agent's configuration and never returned by a tool. If it reaches a chat, revoke it.
• Requests that carry the key never follow redirects. Address overrides must use https (plain http only for this machine) and point at junis.ai, a subdomain of it or this machine, unless JUNIS_ALLOW_CUSTOM_HOST=1 is also set.
• Deletions, unlinks, whole-text and list replacements, revealing stored secrets, switching the web app's organization, Knowledge calls that replace data, and action runs (which act in your own systems) stop for your answer. In Claude Code the plugin's hook turns them into a permission prompt. In lazy mode they return needs_confirmation until the agent calls again with confirmed=true. In full mode without the plugin (Claude Code set up by install.sh, or --full), platform tools rely on their destructiveHint and your agent's approval prompt; junis_api and secret-revealing reads still return needs_confirmation. From the command line without a terminal, such commands exit with code 2 until repeated with --yes.

SECURITY.md lists everything the kit stores, how to revoke a key and how to report a vulnerability.

Limits

• The agent sets confirmed=true itself, so the kit cannot prove that you answered.
• The hook recognizes kit commands by name and shape. A wrapper script that calls the kit from inside is not caught.
• Treating text from Junis as data is a rule for the agent, not a technical barrier.
• A service key you hand the agent passes through the conversation and your model provider. Enter it at https://junis.ai/team/mcp-skills instead.
• claude.ai web and mobile connectors accept only OAuth, which the hosted server does not offer, so the kit works in desktop and command-line agents.
• The kit cannot approve an action, disable one or approve a waiting action run. A person does that in the web app; an action the kit drafted does nothing until then.
• Runs started by the kit do not stream. cancel_run stops one after the current step; that step finishes and is billed.
• The kit sees no prices or spending. The payer reads them at https://junis.ai/usage.
• For work longer than about 20 minutes, the skill uses a one-off schedule.

Command line

The same features work from a shell, with JSON results on stdout. Run it as node ~/.junis-kit/skills/junis/scripts/junis.mjs <command>. A Claude Code plugin install keeps the file under ~/.claude/plugins/cache/junis-kit/ instead, and a skill-only install keeps it in the agent's skills folder.

junis.mjs status | login --browser | login --from-clipboard | logout | org use <name> | doctor
junis.mjs tools <words> | tools describe <tool> | tools call <tool> key=value instruction=@file
junis.mjs api <METHOD> <path> [--json '{...}']
junis.mjs run orchestrator "<message>" | run <agent> "<message>" | session wait <id> | session messages <id>
junis.mjs lint <file> | snapshot [--org id] | snapshot diff <older> <newer>
junis.mjs install | uninstall | mcp [--lazy]

--help lists every flag. A run that outlasts the wait returns a session id; the kit never sends a run twice.

junis.mjs doctor, or asking your agent to check the Junis connection, checks Node, the bridge, the hosted MCP server, the key, the organization and each agent's registration, and names the fix for each failure. Error messages and their fixes are in skills/junis/references/troubleshooting.md. Steps only you can do in the web app (plans, OAuth apps, publishing, team roles) are in skills/junis/references/user-steps.md.

Testing

Latest full eval run, November 4, 2026, on the skill as it was that day and the 20 cases that existed then. Later changes have not been re-measured, and the five Actions cases added since have not been run.

Results are not committed (evals/results/ is ignored).

The handshake tests drive the bridge over stdio in both MCP protocol eras; they do not start the agents. Whether agents other than Claude Code follow the skill has not been measured. npm test runs against an in-process mock of the platform; nothing in the tests or evals touches a real Junis organization.

Contributing

Issues and pull requests are welcome. Development rules are in AGENTS.md, the design rationale in docs/ARCHITECTURE.md and the research summary in docs/METHODOLOGY.md. Run npm run check before you open a pull request; it checks that every manifest matches kit.json and runs the test suites. A change to the skill's behaviour comes with an eval case. Change README.md and README.ko.md together.

Every commit on main reaches plugin users. For a fixed version, install with install.sh and a tag (JUNIS_KIT_REF). To release, run node scripts/build-manifests.mjs --set-version <x.y.z> and npm run check, then tag the commit v<x.y.z>.

License

Apache-2.0: see LICENSE and NOTICE. Under its terms, anyone may use, modify and redistribute the kit, and the license includes a patent grant. It grants no right to use the Junis name or marks beyond describing the kit's origin.

Contact

• Bugs and questions about the kit: https://github.com/TOOEASY-crew/junis-kit/issues
• Security issues: follow SECURITY.md; do not open a public issue.
• Account, billing and platform questions: contact@junis.ai

Junis is built by TOOEASY Inc.