Skip to content
Early previewRuns entirely on your machine

Many agents. One direction.

One local table for every coding agent in your repo: who owns what, who asked what, and what is actually done.

Get started See how it works
Claude Code
Codex
You
OpenCode
Shared state
Illustrative sequence. Integration status varies

Sits alongside the agents you already run

  • Claude Code
  • Codex
  • Any MCP client
  • OpenCode
  • Qwen Code
  • Junie
  • Antigravity
  • Muse
Integration status

Run three agents on one repo and you become the referee. Two rewrite the same file. One asks a question nobody hears. Another says done, and nobody checked. Majlis gives them a table to sit at.

مجلسmajlis: the room where people sit together to talk things through and decide.

How it works

Four moments where parallel agents collide.

Pick one. See what each agent did, and what Majlis kept for later.

Ownership

T-14

Ship the auth flow

Clear scope before the next edit.

  1. Agent:
    Claude Code claims the work

    Explicit ownership of src/auth/**

    Claimed
  2. Majlis:
    Codex sees the overlap

    Another session plans to touch the same scope.

    Advisory
  3. Majlis:
    Keep the decision with the task

    Use the existing session model.

    Accepted decision
OutcomeOverlap advisory

Default warn policy records the overlap and lets work continue. Strict enforcement needs explicit operator configuration and supported hooks.

Explore ownership

Questions

T-14

Ship the auth flow

Ask the session that has the context.

  1. Agent:
    Claude Code asks Codex

    “Which response should an expired session return?”

    Addressed question
  2. Agent:
    Codex answers in the task thread

    “Use the existing unauthorized response.”

    Reply retained
  3. Majlis:
    The answer stays with the work

    Available for the requesting session to retrieve.

    In inbox
OutcomeAnswer available

A retained answer can be retrieved from the inbox. Availability is separate from confirmation that the agent read it.

Explore questions

Review

T-14

Ship the auth flow

Passing checks is one part of finishing.

  1. Agent:
    Verification results are recorded

    The example change passes its checks.

    Checks passed
  2. Majlis:
    The exact artifact is submitted

    Review stays attached to this version of the work.

    Submitted
  3. Person:
    The assigned reviewer decides

    Only authorized acceptance advances the task.

    Review pending
OutcomeAwaiting assigned reviewer

Submitted work remains in review. An authorized reviewer must accept the same artifact before it counts as complete.

Explore review

Recovery

T-14

Ship the auth flow

Pick up from evidence, not assumptions.

  1. Agent:
    A session disappears

    Inspect any retained checkpoint and recorded scope.

    Session unavailable
  2. Person:
    Inspect the retained evidence

    Unknown writer outcomes stay visibly unknown.

    Outcome unknown
  3. Person:
    Preview a reviewed takeover

    Inspect the proposed fresh worktree before recovery.

    Preview only
OutcomeRecovery preview

Review the available evidence and recovery plan before authorizing takeover. A preview has not recovered or resumed the work.

Explore recovery

Interactive example with synthetic data

Explore the coordination model

Built for the way agents actually work.

Starts itself.

Run the daemon once. New Claude Code and Codex sessions enroll on their own, under an audited warn policy.

Decisions outlive the session.

Accepted decisions stay with the project. The next session is briefed on the ones that matter.

Your approvals stay yours.

Agent messages are data. They never approve a command or click a harness permission.

One binary. Your machine.

No account, no cloud, no CDN. State stays in ~/.majlis.

binary
1
MCP tools
52
agent adapters
7
accounts
0

Every agent. One board.

A live dashboard ships inside the binary. No Node runtime, no CDN, no sign-in.

majlis ui
Majlis dashboard Tasks board with columns for Open, In progress, Blocked, In review, Done, Orphaned and Cancelled, each card showing its task number, owner and checks.
Majlis dashboard Needs you list: an approval, a proposed decision, a review, a question, a contested decision and a recovery, each tagged by kind and whether it blocks work.
Majlis dashboard Channels view with a general channel where one agent asks another a question and receives an answered reply in a thread.
Demo workspace with fixture data
Dashboard and terminal views

Three commands. Then your usual workflow.

Experimental and unreleased: you build it from source today.

  1. Build

    Build the binary and dashboard from your checkout. The repository is not public yet.

    make build-majlis
    Prerequisites

  2. Start the daemon

    It provisions a local operator and wires up the agents it detects.

    ./bin/majlis daemon

  3. Keep working

    Carry on in Claude Code, Codex or any MCP client. Open the board whenever you like.

    majlis uiCheck your integration path

Less refereeing.
More building.

Get started Read the docs