Odel
Consensus

Consensus

@masterblaster14PythonMITUpdated Yesterday

Coordination for AI coding agents: declare plans, catch design conflicts early, share team memory.

View on GitHubSign in to connect
Server endpointStreamable HTTPAPI keyProbed
https://consensus-production-aed6.up.railway.app/mcp

This is the third-party server itself — Odel doesn't run it. Hitting this URL directly talks straight to the upstream server with no auth or proxying. Connect through Odel to front it with managed auth.

Consensus

Coordination for teams whose developers each run an AI coding agent on the same codebase.

Before an agent writes code, it declares its plan. Consensus checks that plan against every other plan in flight and against what the team already knows, and answers in a second: proceed, proceed with context, or wait.

Why

AI coding agents are fast and they work alone. On a team that produces two problems git cannot see:

  • Design conflicts in files that never touch. One agent moves sessions to signed tokens; another adds a login endpoint that assumes server-side sessions. No merge conflict. Two incompatible designs discovered days later in review.
  • Every agent starts from zero. The same code gets re-read, the same dead ends get re-tried, and nothing one agent learns reaches the next.

What it does

  • Declare before writing. Plans are compared on what they mean, not which files they touch. Conflicts are caught before the code exists.
  • Shared memory. Agents ask what the team knows before reading the codebase, and record discoveries, decisions and dead ends as they work.
  • Human rulings that compound. When two plans genuinely conflict, a person decides once. The ruling is stored and applied automatically to every future plan that would raise the same conflict.
  • Clean handoffs. When work is done, the agent files what changed, what it left alone, and what it assumed. That becomes the pull request.

Consensus never reads, writes, executes or merges code.

How it works

  1. One language-model call extracts a plan's stance: the concepts it touches and its positions on error handling, authentication, data access and API shape. Unaddressed axes stay empty; nothing is guessed.
  2. The nearest open plans and memory entries are retrieved by vector search.
  3. Plans are compared deterministically. Shared concept plus disagreeing positions is a clash. Same inputs, same answer, every input logged.
  4. A hard clash is checked against prior rulings before any human is asked.

Full detail: How it works.

Connecting an agent

Consensus is an MCP server. Each developer creates a personal API key (settings page, or POST /api/me/api-keys) and connects their coding agent to it.

Claude Code, with the guardrail (recommended). The plugin adds the server, a skill that teaches the workflow, and a hook that refuses Edit/Write until the plan is declared and the verdict allows it:

export CONSENSUS_API_KEY="csk_..."          # CONSENSUS_URL defaults to the hosted instance
claude plugin marketplace add masterblaster14/Consensus
claude plugin install consensus@consensus

Claude Code, server only:

claude mcp add --transport http consensus https://consensus-production-aed6.up.railway.app/mcp --header "Authorization: Bearer csk_..."

Cursor, Windsurf, any MCP client. Add to the client's MCP config:

{ "mcpServers": { "consensus": { "url": "https://consensus-production-aed6.up.railway.app/mcp", "headers": { "Authorization": "Bearer csk_..." } } } }

Listed in the MCP Registry as io.github.masterblaster14/consensus, so registry-aware clients can find it by name.

Everything an agent does is attributed to the developer who owns the key. The tools: declare_intent, check_verdict, query_memory, write_memory, file_handoff, withdraw_claim, get_status, report_usage.

Teams and integrations

  • Sign in with GitHub or an email link. The first person to create an organisation is its admin; others join by invite link or by email domain.
  • One shared live board per repository.
  • GitHub: handoffs open pull requests, rulings become PR comments, merged PRs retire their plans.
  • Notion: tasks sync in, decisions and rulings mirror out.

Hosted instance: https://consensus-production-aed6.up.railway.app (API docs at /docs, live board at /board).

Running it

Python 3.11, PostgreSQL with pgvector, Redis. Docker Compose for the databases. Offline mode needs no API keys.

docker compose up -d
.venv/Scripts/python -m alembic upgrade head
.venv/Scripts/python -m scripts.seed_demo
.venv/Scripts/python -m uvicorn app.main:app --port 8000

Or the whole stack in containers: docker compose --profile full up -d --build.

Setup details, configuration and the full API: Backend reference. Hosting it: Deploying (Render blueprint, Railway, Fly, or any Docker host).

License

MIT. See LICENSE.

Documentation