The Governance LayerCivicAide
The agent harness is becoming an industry standard. Anthropic, OpenAI, LangChain, and the open-source community have converged on five core components: context management, tool access, state persistence, execution orchestration, and model-agnostic design. Within a year or two, this base infrastructure will be table stakes. What none of them are building is the accountability layer that local government requires before any of it can be deployed in public service.
That's what CivicAide is: an open-source governance layer built on top of the best harness infrastructure the industry offers. Decision traceability so every AI-assisted output traces back to its inputs and reasoning. Public records compliance so AI interactions meet FOIA requirements. A verifiability layer that helps surface which tasks are good candidates for automation and which still need human judgment, based on structured criteria. Approval workflows that match how local government actually makes decisions, where staff approve operational tasks and elected officials approve policy-adjacent outputs.
This isn't about restricting AI. It's about making it deployable. An elected official can say yes to AI when they can show constituents that every output is traceable, every decision point is logged, and the city controls the infrastructure. The governance layer is what turns a capable technology into a trustworthy public tool.
CivicAide started as a Coda doc for one part-time councilmember processing 200-page council packets. As frontier models and agent frameworks matured, so did the ambition, from personal toolkit to open-source governance infrastructure for the entire municipal organization. Currently built on Anthropic's infrastructure, designed so cities are never locked to a single provider.
The Origin
Built from actual workflow, not a theory of it.
CivicAide was originally built for part-time elected officials who are short on time due to day jobs, families, and the reality that local governance is a second (or third) shift. The first version was a Coda doc connected to various LLMs, using prompt chaining to process the 200-page council packets that land every week.
As frontier models improved and agent frameworks matured, the scope grew. Not because we went looking for problems, but because the tools became capable enough to address the ones we were already drowning in.
Not everything in CivicAide is AI. Automation is sufficient where it fits. The goal is capacity, not complexity.
PolicyAide
A multi-agent system for policy research: enter a topic, receive a professional-grade policy brief with citations, argument analysis, and compliance checking. Built on Anthropic's Agent SDK, inspired by Google's AI CoScientist paper. PolicyAide is the furthest along of CivicWork's tools and demonstrates what becomes possible when governance infrastructure is part of the design from the start.
The core insight: policy decisions are non-verifiable at the decision level (that's what democracy is for), but research quality itself can be verified through adversarial testing. PolicyAide stress-tests the reasoning before it reaches the decision-maker. Council makes the decision; the system makes sure the analysis is rigorous.
Specialized Agents
Supervisor
ResearchAnalyzes the research scope, determines data needs, and creates a structured research plan with exploration areas and success criteria.
External Data
ResearchWeb-grounded research across six categories: demographics, economics, legislation, comparable jurisdictions, news, and academic literature.
Generation
ResearchProduces diverse policy hypotheses spanning status quo to transformative approaches, grounded in the research data and uploaded documents.
Debate
TournamentHead-to-head policy evaluation. Builds cases for and against each hypothesis pair, delivers confidence-weighted verdicts that feed the ELO system.
Evolution
TournamentRefines losing hypotheses with additional evidence and reframing between tournament rounds. Validates that improvements don't introduce bias.
Proximity Check
QualityValidates novelty and overlap across the final policy options. Recommends merges for redundant proposals, ensures a diverse final set.
Meta-Review
SynthesisSynthesizes all agent outputs into a professional seven-section policy brief with citations, compliance checks, and bias detection.
Citation Verification
QualityVerifies every citation URL through cross-referencing and targeted web search. Labels each source as verified or unverified, and never fabricates URLs.
The Pipeline
Each research session runs through a tournament pipeline: hypotheses are generated, then compete head-to-head in ELO-rated debate rounds. Weak proposals evolve between rounds. The system stops when ratings converge or the round cap is reached.
Research
Supervisor→
External Data→
Generation
Tournament Loop
Repeats until ELO ratings converge
Debate
Head-to-head hypothesis evaluation
ELO Update
Bayesian rating adjustment
Evolution
Refine losing hypotheses
Convergence
Check if ratings stabilized
Debate → ELO → Evolution → Convergence → Debate …
Synthesis
Proximity Check→
Meta-Review→
Citation Verification
The Output
Every session produces a seven-section policy brief: the format a policy analyst would write for a council member, not a chatbot summary. All claims are cited. Citations are independently verified.
1Executive Summary
2Background & Legislative Context
3Comparable Jurisdiction Analysis
4Stakeholder Impact Assessment
5Fiscal & Economic Analysis
6Implementation Considerations
7Key Questions for Deliberation
Each brief includes minimum five comparable jurisdictions, equity dimensions, fiscal confidence levels, and five to eight decision-forcing questions for council deliberation. No recommendations. The system informs the decision, it doesn't make it.
The Origin: Elected Official Tools
CivicAide is designed so that AI tools reach staff before individual elected officials, building institutional capacity rather than creating information asymmetries.
That said, CivicAide started as a complete toolkit for one part-time councilmember. These tools proved the concept. The day-to-day workflow of elected officials was the first problem space where the approach was validated.
Orientation Aide
Essential onboarding for new elected officials: Open Meetings Act, Robert's Rules, ethical finance compliance, and a comprehensive transition guide.
Meeting Notes Aide
Create and manage meeting notes with automatic extraction of action items and newsletter content into organized tables for follow-up.
Action Item Aide
Task organization with automated reminders to ensure timely completion of commitments made in meetings and constituent interactions.
Newsletter & Email Aide
Draft and send professional newsletters and constituent emails with built-in Open Meetings Act compliance checking.
Project Status Aide
Track municipal project progress without complex project management software. Built for the way local government actually works.
Network Aide
A directory of local officials and taxing bodies for efficient collaboration with key stakeholders in your community.
Staff First, Then the Whole Organization
The elected official tools proved the concept, but CivicAide is designed for organizational deployment, starting with staff. The underlying data ingestion pipeline serves the entire municipal organization: city manager's office, community development, public works, and beyond.
With the arrival of organizational AI plugins, some of these may be deployments and configurations of shared tooling rather than standalone applications, meeting municipalities where their teams already work.
Data Layer & Agent Readiness
Municipal data is scattered across vendor platforms that don't talk to each other. Staff spend significant time manually looking things up across multiple systems. AI tools can't help if they can't access the data.
Working with U.S. Digital Response in Elgin, we're mapping where municipal data lives, what's locked behind vendors, and what can be freed using open-source tools. Today, answering “what's the status of the water main project on Highland Avenue” requires checking three systems and calling someone back. In an agent-ready environment, an AI assistant pulls that answer in seconds because the data is structured and accessible. That's not replacing the public works staffer. It's giving them back the 15 minutes they spent toggling between systems.
The WebMCP integration for municipal platforms (Legistar, Municode, public meeting archives) is built and being published to GitHub. These tools use open standards to let AI systems read from the platforms municipalities already use. No custom APIs required.
Open-Source Tooling
CivicWork builds for model independence; the tools are designed so municipalities aren't locked into any single AI provider, the same way they shouldn't be locked into any single data vendor. The Coda MCP Server , a 34-tool integration layer for AI systems, is already used by Coda's engineering team to inform their official implementation.
The Municipal Plugin for Claude Code is live and open source. WebMCP integrations for municipal platforms are shipping soon. The governance frameworks, agent architecture, and data layer work regardless of which model powers them.
In development: the Municipal Plugin is adding a vendor contract analysis workflow. Upload a contract. See what you're paying for in plain language. Understand what could be built with open tools. It's designed to make the cost of vendor dependency legible, not to attack vendors, but to give municipalities the clarity to make informed decisions about their own infrastructure.
See the frameworks in action.
The Governance Assessment applies the Verifiability Framework and Trust Stack to your city's AI use case. Try it, or reach out directly.