A task manager your AI agents keep themselves.

English · Русский

You work with Claude Code or Codex as usual and never mention a task manager. The agent creates the main tasks,
takes them into work, marks what blocks it, completes them with a result and drops what you abandoned. After a
session Conus shows what is in progress and what got done: who changed what, when and in which session, with the
commits.

Conus tracks main tasks only: features, notable fixes, important bugs, decisions that lead to separate work. Small
edits and the steps of an agent's plan never become tasks; they end up in the result of the main task.

How it works

"The agent keeps the tasks itself" takes three levels, each with its own mechanism:

flowchart LR
  A[Claude Code / Codex / Gemini CLI] -- MCP over stdio --> B[conus mcp]
  A -- hooks --> C[conus hook]
  B -- HTTPS --> S[(Conus server + Postgres)]
  C -- HTTPS, outbox when offline --> S
  W[Web app] --> S

On every agent machine conus-cli runs the agent's hooks and the conus mcp bridge. Hooks report facts (edited
files, commits, test runs) and hand the agent a short snapshot of its open tasks at the start of a session and on
every prompt. At the end of a turn the server checks the facts against the board and, when a main task was
clearly left behind, asks the agent to update it. When the server is down, nothing stops: events wait in a local
outbox and the agent keeps working.

Quick start

1. Run the server

You need Docker with Compose v2.

curl -fsSL https://raw.githubusercontent.com/Pandusss/conus/main/deploy/install.sh | sh

The installer puts compose.yml and a generated .env (a random database password, a setup code) into ./conus,
starts the server and prints its address and the setup code. It asks one question: a domain for HTTPS.

• No domain: Conus runs on http://localhost:8080 of this machine. Good for trying it and for agents on the same
  machine.
• A domain whose A record points at the server, with ports 80 and 443 free: Caddy gets a Let's Encrypt
  certificate and Conus is on https://<domain>.
• Your own reverse proxy: answer with no domain, then point the proxy at 127.0.0.1:8080 and set
  CONUS_PUBLIC_URL in .env to the public address.

Open the address, enter the setup code and choose the owner password.

2. Connect your agents

In the web app open Agents → Connect agent, name the agent and copy the two commands. On the agent's machine
(Node.js 22 or newer):

npm install -g conus-cli
conus setup https://conus.example.com conus_xxxxxxxx

No access to the npm registry? Every release also carries the package:
npm install -g --allow-remote=all https://github.com/Pandusss/conus/releases/latest/download/conus-cli.tgz (--allow-remote=all lets npm 12 and newer install from a URL; older npm ignores it).

conus setup logs in, then connects every agent it finds: the Conus plugin in Claude Code (through the
claude CLI; without it, it prints the two /plugin commands to run in Claude Code), the MCP server, hooks and
skill in Codex, the MCP server in Gemini CLI. It offers to add a short Conus rule to each agent's global
instructions and finishes with conus doctor. Restart the agents afterwards. In Codex, approve the hooks once
with /hooks.

To install only the Claude Code plugin by hand (after conus login <url> <token>), run in Claude Code:

/plugin install conus --marketplace Pandusss/conus

Any other MCP client: add a stdio server that runs conus mcp, and add the protocol block to the repository with
conus init --agents-md.

That is all: from now on the agents keep the board themselves.

Privacy

Only this leaves an agent's machine:

• paths of changed files, relative to the repository root;
• sha and subject of commits;
• the git remote address (without login or password), the branch, the HEAD sha, the repository folder name and
  a hash of its path;
• the working directory relative to the repository root;
• the agent's model name;
• command lines of test runs (up to 300 characters, simple commands only — never from scripts with a heredoc, and
  for PowerShell never with a here-string, backtick line continuation or -EncodedCommand);
• excerpts of the prompt and the answer, up to 2000 characters;
• titles and texts of tasks written by the agent;
• service data: the agent's session id, the names of the Conus tools called, task refs.

File contents and diffs are never sent. Prompt and answer excerpts can be turned off; the server then works from
facts only (commits and files):

conus config set send_text false

The token lives in ~/.conus/config.json, readable by its owner only; the server stores only its SHA-256 hash, and
you can revoke it in the web app. The optional judge (see below) sends the turn's facts and excerpts to the
Anthropic API.

To stop tracking in one repository, run conus disable in it (conus enable turns it back on). Folders outside git
are never tracked.

Configuration

Server (.env next to compose.yml)

Update: run the installer again, or docker compose pull && docker compose up -d in the conus folder.
Back up: docker compose exec db pg_dump -U conus conus > conus.sql.
Forgot the owner password: docker compose exec app node dist/admin.mjs reset-password.

Agent machines

conus setup writes ~/.conus/config.json. CONUS_URL and CONUS_TOKEN override it (handy in CI),
CONUS_CONFIG_DIR moves it. Messages are in English; conus config set lang ru (or CONUS_LANG=ru) switches
them to Russian.

Troubleshooting

• Run conus doctor: it checks the configuration, the server, the token, that conus is on PATH, each agent and a
  test session in the current repository. ✓ is fine, ! is a warning, ✗ is an error (exit code 1).
• Hooks never interrupt the agent and stay silent on errors; they log to ~/.conus/logs/hook.log.
• ✗ token: the token was revoked or mistyped. Issue a new one in the web app and run conus setup again.
• ! path: agents start conus from hooks; add npm's global bin directory (npm prefix -g) to PATH, for desktop
  apps too.
• ! codex with a hint about /hooks: open Codex and approve the hooks.
• The server is down: work goes on; events collect in ~/.conus/outbox and are sent on the next successful call.

Contributing

Questions and ideas go to Discussions; bugs to issues; pull requests
are welcome, see CONTRIBUTING.md. Report vulnerabilities privately:
SECURITY.md.

License

MIT