n3rv
Invisible engineering infrastructure for opencode AI agents: memory, skills, and subagent dispatch, bootstrapped with a single command.
- Source repo
- juanmanueldaza/n3rv
- Stars
- ★ 11
- Last updated
- 1mo ago
- License
- Apache-2.0
- Primary language
- Python
- FA score
- 49/100 · Major gaps
At a glance
- How it runs
- Works with
- Universal · cross-platformOpenAI API
- Cost
- Free software; you pay for model usage
- Setup effort
- Medium · a few setup steps
- You'll need
- Typical use
- opencode users who want to scaffold a complete workspace (AGENTS.md, skills, plugins, git hooks) for a new project in one command
- Not a fit if
- Teams that want a hosted, turnkey agent and do not use opencode
- Users with no DashScope/OpenAI keys and no capacity to run local Ollama/vLLM models
- Website
- n3rv.daza.ar ↗ set on GitHub
- Maintenance
- Commits in the last 6 monthslast commit 2026-10-13latest release v0.1.2 (2026-10-13)
- Source review
- 49/100 · Major gaps
What does this agent do, and when should you use it?
n3rv is a runtime library and config generator for opencode agent workspaces, maintained by Juan Manuel Daza. Running n3rv init generates AGENTS.md, opencode.json, skill/command/plugin directories, and .n3rv/ memory and hub configuration. Runtime capabilities include a provider abstraction (Qwen/DashScope, OpenAI, local Ollama/vLLM), an HTTP-based A2A hub for cross-process task delegation, and MAGI dual-store memory (ChromaDB vector search plus SQLite relations). It also ships two MCP stdio servers, n3rv-memory and n3rv-hub, consumable by any MCP client. Licensed under Apache-2.0, it positions itself as the implementation counterpart to the satellite protocol specs in the companion repo reverberage/hub.
After installation, n3rv init uses Jinja2 templates to generate AGENTS.md, opencode.json, .opencode/ (skills, agents, commands, plugins), .n3rv/ (memory, hub config), and .githooks/ in your project. At runtime, the ModelProvider Protocol and get_provider() factory resolve the LLM backend from the N3RV_PROVIDER environment variable (qwen/openai/local), so satellites never hardcode model IDs or API keys. The A2A hub, started with n3rv hub start, listens on 127.0.0.1:19820, routes cross-process agent tasks by skill_id, and persists results to JSON files. MAGI memory writes SDD artifacts, session summaries, and agent verdicts into ChromaDB (semantic search) and SQLite (relations), manageable via n3rv memory list|search|prune|stats. The MCP servers n3rv-memory and n3rv-hub expose memory operations and task delegation over stdio. n3rv daemon manages a background daemon, and n3rv org manages multi-project workspaces with shared skills and GitHub branch protection.
- opencode users who want to scaffold a complete workspace (AGENTS.md, skills, plugins, git hooks) for a new project in one command
- Teams delegating tasks across multiple opencode processes via the A2A hub, routed by skill_id
- Developers who want long-term semantic memory for agents, retrieving past session summaries and verdicts from ChromaDB
- Qwen-on-DashScope users who want satellites to resolve providers at runtime without hardcoding model IDs or keys
- MCP client users who want stdio access to memory retrieval or task delegation
- Organizations maintaining multiple repos that want shared skills and automated GitHub branch protection
Key setup facts
Pulled from the install and usage notes below. The project README has the current values.
- Install / run
pip install git+https://github.com/juanmanueldaza/n3rv.gitpip install -e ".[dev]"
- Environment variables
DASHSCOPE_API_KEYN3RV_DEFAULT_MODEL
How do you install or deploy this agent?
Install directly from GitHub with pip, then initialize inside your project:
bash
pip install git+https://github.com/juanmanueldaza/n3rv.gitTo run the test suite, install dev dependencies:
bash
pip install -e ".[dev]"pytest
How do you use this agent?
Initialize a workspace in your project directory (optionally choosing a stack):
bash
cd your-project
n3rv init [--stack python|node|go|generic] [--force]Set the required environment variables (default provider is qwen):
bash
export DASHSCOPE_API_KEY=...
export N3RV_PROVIDER=qwen
export N3RV_DEFAULT_MODEL=qwen3-coder-plusStart the A2A hub, manage the daemon and memory:
bash
n3rv hub start
n3rv daemon install|start|stop|status|enable|logs
n3rv memory list|search|prune|stats
n3rv org init|list|add|remove|protectWhat are this agent's strengths and limitations?
- A single
n3rv initgenerates a complete opencode workspace (AGENTS.md, skills, plugins, git hooks), eliminating manual setup - Provider abstraction supports Qwen, OpenAI, and local Ollama/vLLM, with runtime resolution preventing vendor lock-in
- A2A hub provides HTTP-based cross-process agent task delegation routed by skill_id with JSON persistence
- MAGI dual-store memory combines ChromaDB semantic search with SQLite relations for session summaries and verdicts
- Built-in MCP stdio servers (n3rv-memory, n3rv-hub) are reusable by any MCP client
- Core workflow targets the opencode ecosystem; teams not using opencode must assess adaptation costs of the generated configs
- Defaults depend on DashScope/Qwen (including a specific default base URL); switching providers requires understanding the env-var system
- MAGI memory requires maintaining dual ChromaDB and SQLite stores, adding runtime and data-management overhead
- Beyond the README there is no standalone deployment documentation or release notes, so deep integration requires reading the code
- The upstream repo (juanmanueldaza/n3rv) is GPL-2.0 while this repo is Apache-2.0; the license difference warrants legal confirmation
How does this agent compare with similar options?
The README explicitly divides work with the companion repo reverberage/hub: hub ships the satellite protocol, specs, and scaffold scripts (contracts), while n3rv ships the runtime engine (providers, A2A hub, memory, daemon, CLI, init). Both are Apache-2.0.
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| n3rv This agent | 49 · Major gaps | CLIFree + model costs | ★ 11 | 1mo ago | Python | OpenAI API |
| FastAgent | 60 · Some gaps | CLIFree + model costs | ★ 4k | today | Python | Codex · OpenAI API · Claude API |
| lcc — Local Context Compiler | 77 · Good | CLIFree | ★ 20 | 2d ago | Python | Claude Code |
| Compartment | 73 · Some gaps | CLIFree | ★ 473 | 3d ago | Python | Codex · Claude Code |
How does FollowAgents rate this agent?
Why each dimension lost points
The README makes local persistence, SQLite/ChromaDB, hub state, environment variables, and the localhost service visible, so data-flow transparency scores 2. Memory defaults to full access including delete/prune; the daemon can be enabled at login; and init --force can overwrite files. The packet does not show confirmation before these operations, detailed permission controls, or a general recovery path, so least privilege, confirmation, external effects, and rollback score 1. API-key variables are documented, but key protection and retention/deletion handling are not shown, so sensitive-data handling scores 1. Dependencies have lower bounds but no upper bounds. A ChromaDB comment records a known CVE and rationale for local mode, yet the declared range still includes the described unfixed release pending a future fix; dependency security scores 1. README and pyproject name maintainers, but publisher identity remains unknown as instructed; attribution scores 2 without affecting other criteria.
The README, CLI, and static test samples provide some corroboration for subsystems, task recovery states, and tool conventions, so self-consistency scores 1. Dependencies have only minimum versions, and the key ChromaDB note still concerns an unfixed release, so dependency availability scores 1. The CLI converge command only prints guidance, contrary to the README claim that it reruns verification and generates tasks; evidence for failure-message quality is limited, so that criterion scores 1.
The README targets opencode engineering workflows and names Python, Node, Go, generic, multi-satellite, and SDD scenarios, so audience and scenarios score 2. The code graph's Python AST scope, MCP tools, and safe memory profile provide some capability boundaries, scoring 2. The sample does not show trigger conditions for the 18 skills and several tool classes, so trigger precision scores 1. Python 3.11+, opencode configuration, systemd user service, and localhost hub settings describe environment fit, scoring 2.
The README organizes subsystems, installation, CLI, MCP, configuration, and linked deployment/tool documentation, so information architecture scores 2; quick installation and scaffolding steps are concrete, so install notes score 2. The naming system is broad, but at least one documented feature does not match its CLI implementation (converge), so naming stability scores 1. CLI and configuration examples, documentation references, and a security policy offer useful guidance, scoring 2 for examples and FAQ. Limitations appear only in scattered notes about Python AST, systemd, experimental background subagents, and the dependency CVE, so known limitations score 1. The LICENSE and package metadata clearly specify Apache-2.0, scoring 3. Version 0.1.2 and a releases link are present, but no changelog is included in the packet, so versioning/changelog scores 1. README and SECURITY.md identify maintainers and vulnerability-reporting channels and targets, so maintenance responsibility scores 2; unverified identity is not itself a deduction.
MCP tools, persistent memory, task routing, a code graph, and execution caching provide structured outputs and integration points for engineering agents, so output usability and marginal value score 2. Running a daemon, writing project configuration, relying on local databases/watchers, and deploying multiple services add operational cost; the packet does not quantify benefits or show cost-reduction evidence, so cost-benefit scores 1.
README claims about tool counts, CLI, environment variables, and layout have partial counterparts in pyproject, CLI, and test samples, making the main claims traceable and scoring 2. Corroboration across sources is incomplete and the converge description conflicts with the CLI, so cross-source corroboration scores 2. Documentation claims, code entry points, and test samples can be distinguished, but some security statements, such as the ChromaDB non-exploitability assessment, remain repository assertions not independently verified by this static packet, so fact/inference separation scores 2.
- Treat the repository's claim that the ChromaDB CVE is not exploitable with local PersistentClient use as an unverified assertion; the dependency range still permits the described unfixed release.
- The README says converge reruns verification and creates remediation tasks, while the shown CLI only prints guidance; verify this workflow before relying on it.
- The default memory profile is full and includes deletion and pruning; inspect project data and service settings before enabling the daemon at login.
Recent releases
Latest entries from the project's GitHub Releases. Notes are quoted from upstream as written, not rewritten or tested by FollowAgents.
-
Release notes excerpt
Patch release fixing the MCP memory_delete tool. - memory_delete raised TypeError: MemoryService.memory_delete() got an unexpected keyword argument 'enforce_profile' on every call. The stray keyword is removed; the service already reads N3RV_MEMORY_PROFILE and gates hard_delete internally. - Added regression tests that invoke the *registered* tool through MCPServer.call_tool. This was the coverage gap: the existing service-level tests passed while the tool itself was 100% broken.View on GitHub ↗ -
Release notes excerpt
What's Changed * build(deps-dev): update uv-build requirement from <0.9.0,>=0.8.11 to >=0.8.11,<0.12.0 by @dependabot[bot] in https://github.com/juanmanueldaza/n3rv/pull/37 * feat(code-graph): CodeGraph-inspired surgical explore with auto-sync by @juanmanueldaza in https://github.com/juanmanueldaza/n3rv/pull/50 * feat: add converge command to n3rv CLI by @juanmanueldaza in https://github.com/juanmanueldaza/n3rv/pull/54 * feat(harness): n3rv-scout/n3rv-sensor specialist subagents + universal exec MCP by @juanmanueldaza in https://github.com/juanmanueldaza/n3rv/pull/64 …View on GitHub ↗
FAQ
Does running n3rv cost money?
Do I have to use opencode?
What are the default provider and model?
qwen, the default model ID is qwen3-coder-plus, and the default base URL points to the DashScope international compatible endpoint; override via N3RV_PROVIDER, N3RV_DEFAULT_MODEL, and related environment variables.How does the A2A hub work?
n3rv hub start, it listens on 127.0.0.1:19820, routes cross-process agent tasks by skill_id, and persists results to JSON files.Is n3rv a satellite?
Is n3rv open source?
Is n3rv safe to use?
Is n3rv still maintained?
Who makes n3rv?
What is n3rv's official website?
Thanks — we use this to improve the page.