Teams now run fleets of coding agents and production-facing MCP tools. Concord gives each task a governed harness — code, deploy, live investigation, bootstrap job, or data product — and records the proof in a single source of truthful intelligence.
✈️ Air-traffic control for code, runtime evidence, data contracts and deploy checks — across Claude · Codex · GeminiConcord follows agent work through every stage of delivery, not only the pull request.
Running parallel coding agents is now a commodity — four in five organizations adopted AI development tools faster than their governance policies could keep up. The unsolved part is proving what happened: before merge, after deploy, and inside production-facing work.
43% of organizations can no longer reliably distinguish agent-written code from human code. When the audit comes, “which agent did what, under which policy, with whose approval” has no answer.
An agent can request, write and act in seconds — collapsing segregation of duties into a single log line. SOC 2, SOX and GxP obligations don't care how the code was generated.
Agents with MCP access read and change live systems — and the evidence of what they saw and did scatters across terminals and chat logs, gone by the time anyone asks.
Figures: GitLab AI Accountability Report (Harris Poll, June 2026). Gartner predicts over 40% of agentic AI projects will be cancelled by 2027 — cost, unclear value and inadequate risk controls; the projects that survive are the ones that can prove control.
The agents are the workforce; Concord is the control plane that coordinates them, gives each a role, and accounts for everything they do — across repos, providers, runtime tools, and proof types.
Every agent works in its own git worktree behind a per-ticket lock — so a fleet runs in parallel, across many repositories, with clear ownership and fewer collisions.
Governed roles take work from plan → build → review → test → land, in either topology: independent agents, or an orchestrator directing sub-agents.
Every lifecycle transition is attributed to a specific agent, under a specific reviewer, in an append-only journal — so multi-agent work can be replayed and audited.
Every unit of agent work runs the same path — and lands in a tamper-evident journal. The proof changes by track, so a live production read is not verified like a code diff.
Worktree + branch + lock per ticket. Many agents, many repos — no collisions.
Tests for code, content checks for pages, receipts for live-MCP work, data contracts for analytics, deploy evidence for runtime work.
Every transition recorded with actor and provenance — a verifiable history of who did what, under whose review.
A read-only web view renders the board, traceability, gates, live-MCP receipts, bootstrap risk, runtime health and timeline.
Concord now governs more than code diffs. It gives each work type the evidence standard it actually needs, while keeping one board, one journal and one recovery model.
Per-ticket worktrees, locks, review cycles, feature proof, contract checks, quality dimensions and source-to-landed commit lineage.
Production-MCP investigations, deploy checks and bootstrap/backfill jobs record scoped, redacted, source-cited evidence instead of relying on "readyz returned 200".
Analytical products certify data contracts, row-count reconciliation, lineage and quality gates. The same journal feeds governed memory: cited decisions, summaries and recall.
Everything below ships in the open-source Community edition except Enterprise visibility. As of v0.1.16.
Each agent works in its own git worktree and branch behind a per-ticket lock, which reduces collisions between parallel agents.
Worktrees separate repositories; they are not a sandbox.
Plan records capture what was asked, built, not built and deferred, with the exact verification commands.
The journal covers Concord lifecycle operations, not every agent tool call.
Import the URS or PRD you already have, with stable IDs and source anchors, then trace each requirement to tickets and evidence.
Business discovery maps an existing repo with confidence levels, so current behaviour is not mistaken for intended behaviour.
Every ticket gets a risk class from R0 to R4 and a computed gate plan; high-risk work without proof cannot enter review.
Five governed tracks, each with its own evidence: development, content, devops, live operations and data.
A read-only cockpit shows the board, gates, requirements, decisions and runtime evidence. It shows commands; it never runs them.
npx create-concord plans the install and applies only the digest you reviewed, for a new project or an existing workspace.
Read-only collection from many repositories into one warehouse, with per-repo throughput, cost, staleness and gate health.
SSO, SIEM export and KMS-backed signing are a defined path, not live controls.
This isn't a framework we hope you'll try — it's the governed execution layer behind our production AI. Our delivery teams run their agent work through it — the “single source of truthful intelligence” we named in 2020, now the system we run on.
Production AI projects governed end-to-end — pharma, clinical, claims, education, hospitality, logistics and data.
Governed, evidence-backed changes shipped — each attributed, gated and journaled.
Peak concurrent agents on a single project — dozens a day — coordinated through per-ticket locks.
Since September 2026 every softsensor.ai change is a Concord ticket: planned, gated in CI and landed through a reviewed pull request.
Across these projects Concord has coordinated 500+ agent identities over 1,200+ sessions and 17,000+ governed events — both orchestrated sub-agents and independent terminal agents, governed identically. Figures reflect Softsensor's own governed delivery; client names are withheld for confidentiality.
Lineage: Concord grew out of AI Agent Setup — our open-source, skill-aware tooling for standing up Claude, Codex and Gemini on any repo (2025, Apache-2.0). First we made agents easy to set up; then we made them accountable.
The question has moved from “did a check run?” to “which agent did what, under which policy — and can you prove it?” Branch protection governs the pull request. Concord governs the agent work before the PR — and turns it into a record an auditor can use.
Required checks, protected branches, CODEOWNERS, approvals. Strong process enforcement — but no record of which agent did what before the PR, no runtime proof after deploy, and no governed memory of the decision trail.
Task locks, per-ticket worktrees, requirement closure, agent provenance, track-specific proof, runtime receipts, fail-closed review and governed memory — across providers, repositories and live engineering workflows.
AI coding agents and MCP-powered runtime work don't remove SDLC, CSV or change-control expectations. Concord is built for the regimes you already answer to — producing review evidence, traceability, runtime receipts and change history as a by-product of the workflow.
Per-change closure, review cycles and an attributable audit trail — the records these regimes ask for.
Traceability and fail-closed, control-mapped export — aligned to the direction of the FDA's Computer Software Assurance guidance (records over screenshots).
Indicative control maps included; the evidence export fails closed if anything required is missing.
We describe Concord as designed for / building toward these regimes — it produces evidence and indicative maps, not a certification, validation, or claim of conformity.
73% of enterprises report AI costs exceeding projections — agent spend is now a governance question too. Governance adds structure, but it removes waste: Concord holds the context, decisions and evidence agents otherwise re-pay to rediscover.
Cheap pre-checks classify already-satisfied work before an agent is ever dispatched.
Decision records, summary tiers, recall and graph context let agents retrieve the relevant slice instead of replaying the whole history.
Evidence depth follows the track — lighter for mechanical edits, stricter for production reads, data products, deploys and bootstrap jobs.
A cost-ledger records spend per change in the same journal that proves the controls — so savings are measured, not estimated. Blended figures will be published from design-partner pilots. Cost figure: FinOps Foundation, State of FinOps 2026.
Concord installs as a coord/ directory beside your code. The installer shows its plan before it writes anything.
# 1 · install into a new project, or overlay an existing repository npx create-concord my-project npx create-concord . --from-existing # 2 · open the read-only cockpit npm run coord-ui
Governance commands need Node.js 22.8 or later. No Node on the machine? The standalone Linux binary installs with concord init .. Full steps in the quickstart.
The repo-local governed workflow is open source. Enterprise packaging adds org-wide collection, policy and deployment patterns for customers who need central command-center evidence.
An open-source (Apache-2.0) governance engine for AI coding agents. Each piece of work gets its own worktree and lock, a risk-sized gate plan and a recorded closure, written to a hash-chained journal.
No. Rulesets govern the pull request; Concord governs the agent work before it and records runtime evidence after it. Its gates call your existing test and build commands.
It is provider-agnostic. Session identity is built in for Claude, Codex, Gemini and Grok, with agents working independently or under an orchestrator.
Yes. npx create-concord . --from-existing plans an overlay you review before anything is written, and business discovery maps the repository without treating current behaviour as intended behaviour.
Yes. Concord coordinates the work that goes through it; an agent or person with local access can act outside it. The design makes that detectable, not impossible, and worktrees separate repositories rather than sandboxing processes.
Org-wide collection and rollups, central re-hash verification with signed bundles, discovery of team boards, and the RBAC broker. Everything else is in the open-source edition.
No. It produces evidence and indicative control maps that your existing programme can use. It is not a certification, validation or conformity claim.
In your repositories. The board, plan records, gate artifacts and journal are files you own; Enterprise reads them and does not write back.
Run coding agents, production-MCP investigations, bootstrap jobs and data products through one governed execution record.