Settled

tests  Python · Slack Bolt · Model Context Protocol · SQLite (hash-chained audit) · MIT

Retrieval tells you what was said. Settled tells you what was decided — and whether it still binds.

Teams run AI agents in Slack now. An agent acts on whatever it reads — even a decision the team already reversed. Settled maintains a decision ledger with epistemic status and, uniquely, exposes it over an MCP server so any agent can ask "is this still binding?" before it acts.

Slack Agent Builder Challenge · New Slack Agent track · Qualifying tech: our own decisions:// MCP server (ungated).

What it does

• Watches channels, detects decisions behind a precision-first confidence gate (uses thread context, catches multiple decisions per message).
• Human ✅ ratifies — only a human makes a decision binding. Wrong "settled" > silence.
• Epistemic lifecycle: proposed → contested → settled → superseded, each anchored to a verbatim quote + permalink.
• Query: /settled <topic>, the assistant (DM/@mention), or the App Home dashboard.
• MCP decisions:// — is_binding(topic), query_decisions(q) + resources, over stdio & HTTP.

What makes it different

Plenty of tools log decisions for a person to read later. Settled is the one an agent can query before acting — the guardrail that keeps humans and agents on what the team actually decided.

	Decision-log tools	RAG / Slack search	Settled
Captures decisions	✅	➖	✅
Knows which one still binds	❌	❌	✅ (epistemic status)
Verbatim quote + permalink anchor	➖	❌	✅
Human ratifies before "settled"	➖	❌	✅
An agent can query it before acting	❌	❌	✅ (MCP decisions://)
Tamper-evident audit trail	❌	❌	✅ (hash chain + verify)

See it live — a coding agent rewrites its own migration before merge:

$ python -m demo.agent_coding_guardrail
│  ✍️  wrote 0007_create_datastore.postgres.sql  (targets Postgres)
│  → mcp.call(decisions://, is_binding, {topic: "datastore"})
│  ← {"binding": true, "statement": "Move primary datastore to Aurora", "permalink": …}
│  ⛔ CONFLICT  my migration targets Postgres — but the team ratified Aurora
│  ✅ rewrote → 0007_create_datastore.aurora.sql  (targets Aurora)
│      - conn: pg-primary.internal   + conn: …cluster.rds.amazonaws.com

Also: python -m bench.benchmark (naive agents act on a stale decision ~60% of the time, Settled 0%) and python -m demo.breadth (a CI gate, an IDE assistant, and a coding agent on one ledger).

Run locally

python3 -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
cp .env.example .env            # fill Slack tokens + OPENROUTER_API_KEY
python -m seed.seed_demo        # seed the ledger (fictional demo data)
python run.py                   # Slack bot (Socket Mode)
python -m mcp_server.server     # MCP server (stdio); SETTLED_MCP_TRANSPORT=http for HTTP

Create the Slack app from slack_manifest.yaml (declares App Home, /settled, the Assistant surface, and events incl. reaction_removed).

Tests: pip install -r requirements-dev.txt && pytest -q — covers the ledger lifecycle, supersede/topic guards, the hash-chained audit log (incl. tamper detection), the classifier gates, and the MCP is_binding / verify_audit_log tools.

Deploy (always-on for judging)

See DEPLOY.md — Railway or Fly, same Dockerfile. Currently live on Railway.

Layout

Path	Role
settled/	config, db, ledger, llm (classify+answer), extraction, blocks, slack_app, agent
mcp_server/server.py	decisions:// MCP server (stdio + streamable-HTTP): is_binding, query_decisions, verify_audit_log
demo/agent_coding_guardrail.py	the wow demo: a coding agent rewrites its own migration after checking Settled
demo/breadth.py · demo/agent_guardrail.py	3 agents on one ledger · the minimal guardrail
bench/benchmark.py	quantified impact: stale-action rate, naive vs Settled
tests/	29 pytest tests (lifecycle, audit-chain tamper detection, MCP, blocks)
seed/ · video/	demo ledger seeders · Remotion video project + architecture PNG

Stack

Python · Slack Bolt · SQLite (hash-chained audit) · Model Context Protocol · OpenRouter (DeepSeek).

Licensed MIT (see LICENSE); third-party components in NOTICE.md. Demo data is fictional.