AGENA
Open-source, multi-tenant agentic AI platform that autonomously analyzes tasks, generates code, reviews quality, and ships pull requests.
- Source repo
- aozyildirim/Agena
- Stars
- ★ 103
- Last updated
- 1mo ago
- License
- MIT
- Primary language
- TypeScript
- FA score
- 56/100 · Major gaps
At a glance
- How it runs
- Works with
- Universal · cross-platformCodex · Claude Code · OpenAI API
- Cost
- Free tier plus a paid hosted plan
- Setup effort
- High · needs real infrastructure
- You'll need
- Typical use
- Engineering teams that want production errors from Sentry / New Relic automatically turned into fix PRs without manual triage.
- Not a fit if
- Teams not hosting code on GitHub or Azure DevOps
- Solo developers unwilling to run Docker, MySQL, Redis and multiple services
- Users who only want inline code completion rather than fully automated PR pipelines
- Website
- agena.dev ↗ set on GitHub
- Maintenance
- Commits in the last 6 monthslast commit 2026-10-09
- Source review
- 56/100 · Major gaps
What does this agent do, and when should you use it?
AGENA (GitHub repo aozyildirim/Agena, MIT license) is a self-hostable agentic AI orchestration platform covering the full task-to-PR loop. Its backend is a monorepo of 6 pip-installable packages (agena-core / models / services / agents / api / worker) that runs a PM → Planner → Developer → Reviewer → Finalizer pipeline via CrewAI role-based agents and a LangGraph five-node state machine, automatically creating branches, commits and pull requests on GitHub or Azure DevOps. Instead of per-token API billing, it reuses your existing Claude Code / Codex CLI subscription through a host CLI bridge, with OpenAI / Gemini API as alternatives. The stack is Python 3.11 + FastAPI + MySQL 8.0 + Redis 7 + optional Qdrant vector memory, with a Next.js 14 dashboard featuring a pixel-art Boss Mode, a visual flow builder and DORA metrics. It supports multi-tenant org isolation, JWT + RBAC, Stripe billing and Free/Pro/Enterprise plans.
Tasks created via the UI or imported from Jira / YouTrack / Azure DevOps / Sentry / New Relic enter a Redis queue (with repo locking), then OrchestrationService drives the LangGraph pipeline: fetch_context (Qdrant vector memory retrieves similar past tasks), analyze (PM agent estimates story points and plans changes), generate_code (Developer agent produces patches), review_code (Reviewer agent reviews), finalize (cleans output, creates branch, commits, opens a PR). A PR-merge webhook auto-resolves the linked Sentry issue. Surrounding modules include the Skill Catalog (the Skill Librarian distills completed tasks into reusable, embedded patterns injected into future agent prompts), History-Grounded Sprint Refinement (per-org Qdrant collections of closed work items for SP estimation), a Runtimes Registry (bridge-server.mjs self-registers and heartbeats every 30s), multi-repo parallel orchestration, task dependencies with auto-queueing, AI Sprint Nudges and the Review Backlog Killer multi-channel nudger, Cross-Source Insights event correlation (with one-click rollback PRs), and Stale Ticket Triage source-side scanning with AI verdicts.
- Engineering teams that want production errors from Sentry / New Relic automatically turned into fix PRs without manual triage.
- Scrum teams on Jira or Azure DevOps who want story-point estimates grounded in their own historical work items.
- Organizations needing a self-hosted, per-org isolated AI coding platform where each team connects its own LLM accounts.
- DevOps leads who want DORA metrics, team health analysis and cross-source correlation (which deploy caused this bug).
- Senior developers who want to codify team coding patterns into a reusable Skill Catalog automatically injected into future task prompts.
How do you install or deploy this agent?
Prerequisites: Docker, Python 3.11, Node.js; for the CLI bridge mode you must be signed in to Claude Code or Codex CLI on the host (claude auth login / codex auth login). Install:
How do you use this agent?
Once started, open http://localhost:3010 (frontend) and http://localhost:8010/docs (Swagger). Create the first user and organization:
What are this agent's strengths and limitations?
- Runs on your existing Claude Code / Codex CLI subscription with no per-token billing — a cost model distinct from metered-API tools.
- Covers the full loop: import work items from Sentry/Jira, auto-create branches and PRs, auto-resolve issues after merge.
- Skill Catalog and vector memory make the platform accumulate codebase-specific patterns over time instead of re-discovering them.
- Monorepo of 6 independent pip packages — individual packages install cleanly, e.g. pip install -e packages/agents.
- Heavy deployment footprint: MySQL, Redis, optional Qdrant, backend, worker, blue/green frontend containers plus a host CLI bridge.
- The CLI bridge must run on the host (not in Docker) and depends on system keychain Claude/Codex auth — a hard constraint in restricted environments.
- PR automation supports only GitHub and Azure DevOps; other hosting platforms would require adaptation.
- No independent performance benchmarks or security audits are provided in the README; validate before production adoption.
How does this agent compare with similar options?
The README explicitly contrasts AGENA with GitHub Copilot: Copilot is AI-assisted coding (you drive, AI suggests line by line), while AGENA is agentic AI coding (AI drives the full task-to-PR pipeline). They are complementary — Copilot for creative work, AGENA for well-defined tasks.
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| AGENA This agent | 56 · Major gaps | Self-hosted serviceFreemium | ★ 103 | 1mo ago | TypeScript | Codex · Claude Code · OpenAI API |
| PandaProbe | 59 · Major gaps | Self-hosted serviceFreemium | ★ 866 | 2d ago | Python | — |
| Agent Graph | 53 · Major gaps | Self-hosted serviceFree + model costs | ★ 239 | 4d ago | Python | — |
| OtoDock | 70 · Some gaps | Self-hosted serviceFreemium | ★ 202 | today | Python | Codex · Claude Code |
How does FollowAgents rate this agent?
Why each dimension lost points
The platform autonomously creates branches, commits, PRs and posts to Slack/Azure/Jira — broad external effects. But confirm-before-apply modals, opt-in defaults (auto-nudge off), cooldown floors and revocable share tokens are documented, so user_confirmation gets 2. least_privilege only 1: PAT proxying and env-var auto-enrollment of runtimes show a wide credential surface with no least-privilege design visible. sensitive_data_handling 2: keychain storage and ADR-0002 explicitly forbids secrets in extra_config. dependency_security 1: versions are pinned but no audit/scan evidence. rollback 1: only the insights module offers a rollback PR. source_attribution 2: AI comments carry a signed 'written by Agena via {model}' line.
Self-consistency is visibly weak: README says 24 alembic migrations and 25 models while CLAUDE.md says 26 and 27; frontend ports drift between 3010/3011/3012 — self_consistency 1. dependency_availability 2: all Python deps exactly pinned, docker-compose and start.sh provided. failure_messages 2: CLI fallback (Claude→Codex) and delivery-aware sent/rate_limited/comment_failed statuses are concretely described.
audience_and_scenarios 2: engineering-team scenarios (incident correlation, review backlog) are well described. capability_boundaries 2: confidence thresholds, <50% skip rule, look-back windows are stated. trigger_precision only 1: the SKILL.md files are prose patterns, not executable skills with defined triggers. environment_fit 2: Docker, host bridge, multi-runtime registry covered.
information_architecture 2: clear package layout, TOC, docs/ADRs, but the README is truncated mid-env-vars and some sections are lists only. install_notes 2: Quick Start plus three CLI install paths, but no verification step. license full 3: MIT text present. versioning_changelog 1: tag-driven release workflow exists but no CHANGELOG file in evidence. maintenance_responsibility 2: SECURITY.md gives 72h response targets, but maintainer is an individual, not a registered organization.
output_usability 2: structured review output (Summary/Findings/Severity/Score), PRs and dashboards are directly usable. marginal_value 2: the skill catalog and history-grounded refinement are a genuine compounding-knowledge design. cost_benefit only 1: 'no API keys, subscription does the work' is claimed, but the deployment cost (MySQL+Redis+Qdrant+6 containers+host bridge) is heavy and unquantified against the benefit.
claim_traceability 1: 'production-ready', 'GPT-5', 'after ~50 tasks carries 20+ patterns' lack supporting data. cross_source_corroboration 1: README and CLAUDE.md contradict each other on migration and model counts; single-source metrics cannot be cross-checked. fact_inference_separation 1: marketing narration (demo gif, 'zero to fully-loaded in under a minute') is interleaved with engineering facts without separation.
- Documentation numbers are inconsistent (migrations 24 vs 26, models 25 vs 27) — verify against actual code before use.
- The platform writes to GitHub/Azure DevOps/Slack/Jira and creates PRs autonomously; review and disable non-essential opt-in switches and audit PAT scope before deployment.
- extra_config is returned unmasked — never store secrets in it.
- Dependencies are pinned but no vulnerability scanning evidence; run dependency audits before production use.
- No CHANGELOG and unverified publisher identity; pin a fixed commit and trial in an isolated environment first.
FAQ
Do I need to buy an OpenAI API key?
Does my code or credentials leave my infrastructure?
What happens when an LLM agent fails?
Can I self-host for free?
Which languages and interfaces are supported?
Is AGENA open source?
Is AGENA safe to use?
Is AGENA still maintained?
Who makes AGENA?
What is AGENA's official website?
Thanks — we use this to improve the page.