The governed pipeline for AI-written code. Bring any coding agent — Weftra governs what it ships: enforced approval gates, a tamper-evident audit trail, and provenance on every change. Self-hosted, single-tenant, Claude-first today; BYOC and SaaS are coming.
Engine-agnostic by design. Run Claude Code, Gemini, Codex, or your own model — Weftra governs the output, not which model wrote it. It doesn't sell a coding agent, so it can neutrally govern anyone's.
Enforced, not advisory
Every change ships through approval gates and a tamper-evident, hash-chained audit ledger — enforced at the merge boundary, not logged after the fact. The one guarantee a logging or observability tool cannot make.
Provenance on every change
A per-feature record of what was requested, who authorized it, which agent implemented it, what gates ran, and what it cost — the AI-change audit trail an auditor can point to, with deterministic run-replay.
Fleet orchestration
One spec, N repos. A persistent cross-project queue with backpressure, per-project policy, and unified cost and status across your whole portfolio — what a single-repo, single-session coding agent structurally cannot do.
Spec-driven pipeline
Requirements → clarify → plan → tasks → implement → validate. The artifact is the spec, not the conversation — repeatable, reviewable output, with separation of duties (the writer of a change never approves it).
Self-hosted today — BYOC and SaaS next
Deploy into infrastructure you already control — your account, your keys, your model endpoints — via Docker/docker-compose today. Packaged multi-cloud BYOC onboarding and a hosted SaaS option are on the roadmap, not shipped yet.
Where this stands today: self-hosted · single-tenant · Claude-first. The queue, the approval, the ledger and the policy stay on your infrastructure, engine-neutral — the agent runs in Weftra's Docker sandbox by default, and a per-run containment tier is not recorded. BYOC packaging, a hosted SaaS option, and Postgres-backed multi-tenant hardening are roadmap chapters — not shipped. New here? Getting Started in 10 Minutes → is the fastest way to see it work, and the trust page → shows the evidence behind these claims, generated from the repo itself. What Weftra is (and isn't) → says plainly what is and is not built.
Weftra is for people who would rather hand off an outcome and review the result than type beside an agent:
Engineering leads and teams who already delegate and review — a contributor that runs your tests itself and can never approve its own change.
Organizations that must prove what AI changed — who asked, which agent and model ran, what passed, who approved, in a tamper-evident record on your infrastructure.
Platform teams with many repositories — one spec across many projects, one queue, one view of status and cost.
Founders and product-minded builders — describe it in plain language; get back a reviewable change.
It's probably not for you yet if you want inline autocomplete, want to skip review entirely, or need a hosted service today. Who Weftra is for →
AI now writes a large share of new code — and it quietly broke software's accountability chain. For human code the chain was implicit: a named engineer opened the PR, a named reviewer approved it. For AI-written code, the accountable human detaches from the diff, and every auditor, regulator, and security team is now asking the same question: "Who authorized this change, and can you prove it?"
Weftra answers it. It isn't a coding agent competing with Claude Code, Copilot, or Cursor — it's the governed pipeline that runs around whichever agent you choose, and makes every change it produces enforced, attributed, and auditable:
Enforced approval gates — full_auto, auto_safe (analyze + clarify), or po_approval (explicit sign-off before work begins), with a submitter who cannot approve their own feature.
A tamper-evident audit ledger — hash-chained run events, exportable to your SIEM, with deterministic run-replay of how any change was produced.
Provenance per change — the request, the spec that governed it, the engine and model that ran, the gates that passed, and the cost — attached to every feature.
Neutral by design — engine-agnostic role routing means you keep your coding agent; Weftra governs its output, not which model does the writing.
That's the difference between "an AI wrote this" (a liability) and "every change the AI produced is governed, attributed, and enforced" (a control you can demonstrate). See Why governed code →.
# 1. Clone and configuregit clone https://github.com/weftra/weftracd weftracp .env.example .env # add ANTHROPIC_API_KEY + GITHUB_TOKEN# 2. Startnpm ci && npm start# 3. Enroll a project (SSH remote — Weftra clones with the host's own SSH key)curl -X POST http://localhost:3100/api/projects \ -H "Authorization: Bearer $ADMIN_API_KEY" \ -H "Content-Type: application/json" \ -d '{"name":"my-project","repo_url":"git@github.com:you/repo.git","autonomy_mode":"auto_safe","auto_create_pr":true}'# 4. Submit a featurecurl -X POST http://localhost:3100/api/features \ -H "Authorization: Bearer $PROJECT_API_KEY" \ -H "Content-Type: application/json" \ -d '{"title":"Add pagination","description":"Add cursor-based pagination to GET /items"}'
See the Getting Started in 10 Minutes guide for the annotated, step-by-step version of this (including the sandbox image build and prerequisites), the Setup guide for full configuration options, and the Tutorial for a worked walkthrough.
