Nextcloud Access Broker

A human-gated MCP server between AI agents and Nextcloud instances. (MCP is the tool-calling interface AI agents use.) The agent holds no standing file-content access: every read and write requires a scoped, time-limited grant that the owner approves through the configured approval gateway, currently a Matrix room. On instances that opt in, the agent can browse file names without a grant. Every operation is audit-logged before it runs.

What it enforces

• Grants: per-task batches of paths with read/write modes, approved by the owner in a private Matrix room (emoji reactions; typed replies for partial approval, custom expiry, revoke, and mass revoke: revoke <instance> / revoke all). Grants expire after 24h by default, request numbers work once, revocation is instant, nothing renews itself, and the agent holds no standing access of any kind.

• The wall: one path-checking function decides every operation (exact component-prefix matching after normalization; traversal, encoding, and confusable attacks fail closed). The trashbin namespace is unreachable, and no permanent delete exists.

• Audit: an append-only, hash-chained log is written before every operation; a failed log write refuses the operation.

• Credentials: Nextcloud app passwords and the Matrix bot token live in config.yaml on the storage host. The agent sees only the MCP endpoint and the agent token. Error messages are scrubbed of credential values, not just keywords.

Related MCP server: Agent File Guardian

Prerequisites

Development: Python 3.11+ (uv venv). Production: Docker and Docker Compose. On Debian/Ubuntu: sudo apt install docker.io docker-compose-v2, then sudo usermod -aG docker $USER (see the user guide for details). You also need one app password per Nextcloud instance you wire, and a Matrix bot account in a private room with you.

Running (development)

uv venv && uv pip install -e ".[dev]"
cp config.yaml.example config.yaml   # fill in room, tokens, instances
.venv/bin/python -m broker.run       # localhost:8765/mcp, bearer auth

Running (production, Docker)

The host decides exposure. docker-compose.yml publishes the broker on one specific interface address of your choosing: a VPN/tunnel interface (recommended when the agent connects over one), a LAN interface, or 127.0.0.1 behind your own reverse proxy with TLS. Change that address on the host at any time; the app never needs rebuilding for it. Do not publish wider than you intend.

# on the Nextcloud host (the storage host):
docker compose up -d --build
# data/ holds the grant store, audit log, and checkpoint; back it up.

Deployment checklist (Gate 7):

1. Verify the trashbin app is enabled on every wired instance (deviation D1 condition: trash routing depends on it), and the files_lock app is enabled where checkout is intended (D5: LOCK returns an actionable 405 without it).

2. Confirm the compose ports address is the interface you intend, and not one you do not: curl to the intended address must return 401 (the auth gate working as intended), and addresses you did not choose must refuse the connection. A timeout means a host firewall (ufw and the like) is dropping the port; allow it scoped to the chosen interface and the agent's address, never as a blanket rule.

3. Mount data/ on persistent storage; the audit log is the record.

4. Fill config.yaml per instance block; personal/work/org blocks are commented out until deliberately wired (they stay absent from the running config until the owner adds them on the host).

Hermes (agent) configuration

On the AI server, add to the profile's mcp_servers (config.yaml via hermes config set, never hand-edit):

mcp_servers:
  nc-broker:
    url: "http://<broker-address>:8765/mcp"
    headers:
      Authorization: "Bearer ${BROKER_AGENT_TOKEN}"
    timeout: 120
    connect_timeout: 60

BROKER_AGENT_TOKEN goes in the Hermes .env; it is the same agent token configured in the broker's config.yaml on the host.

Tools

Agent surface (/mcp, agent token): request_access (batch: instance, reason, items[{path, mode}]), check_access, list_instances (configured instance names + discovery flags; no URLs or credentials — D6.5), list (free under discovery), move, trash, mkdir. The content tools (read, write) are not registered on this surface, so an agent can neither see nor call them (D5: file content never transits the LLM context window).

Transfer surface (/transfer, transfer token, CLI only): check_access, read, write, checkout, checkin. These are spoken by the broker CLI (broker fetch/push/checkout/checkin), which stages files on the client host and hands the agent a local path plus a manifest (sha256, size, lock token). The agent then works on the local copy with normal file tools.

Sync model (D5): read-only work uses on-demand fetch with a manifest freshness check. Read-write work uses checkout/checkin over Nextcloud files_lock: a checked-out file is write-refused on every access path (web UI, sync clients, other agents) until checkin spends the lock token. An abandoned checkout is hard to miss: it shows in the file list and the owner can override it, where an abandoned sync would sit unnoticed.

Result envelope: ok / refused (the wall; do not retry) / error (infrastructure; report it) / pending (wait for the human decision). CLI exit codes: 0 ok, 1 local/usage, 2 infrastructure (retry later), 3 wall-refused (request access, never retry as-is), 4 verification failure (treat as corrupt), 5 clobber refused.

Failure modes (all fail closed)

• Broker down: no access, state preserved (grants resume on restart; pending requests are discarded; silence grants nothing).

• Nextcloud down: clean errors; grants unchanged.

• Approval plane down: no NEW requests; existing grants continue.

• Audit write failure: operation refused.

Testing

.venv/bin/pytest              # 362 tests
.venv/bin/ruff check broker tests review
.venv/bin/python review/attack_battery.py   # fresh wall attacks
.venv/bin/python review/r2_state_probes.py  # store-state probes

Documentation

• Architecture: components, the wall, the grant lifecycle, the audit chain, and the threat model in brief.

• User guide: setup, the approval protocol, running, and troubleshooting.

• Cutover runbook: step-by-step first deployment, with a verification command at every step.

• Security policy: what this defends against, and what it explicitly does not.

• Contributing: the development discipline.

License: GPL-3.0 (see LICENSE).

AI usage disclosure

This software was developed with AI assistance under human direction; the maintainer reviewed, tested, and is accountable for all code (ai-assisted in the emerging W3C/OCaml disclosure vocabulary). The primary implementation model was GLM (glm-5.3), with glm-5.3-flash as the delegated sub-agent model for parallel review and auxiliary tasks. Portions of the implementation, tests, and documentation are AI-generated; every contribution passed through the verification pipeline before release: a 405-test suite with injected-clock determinism, adversarial review with runnable attack scripts (review/), and a secret-scrubbing publish audit.

This is security-sensitive infrastructure for human-gated file access. If you deploy it, do so on the same terms the maintainer does: read the threat model (SECURITY.md), run the attack battery yourself, and treat the approval room, tokens, and audit log as part of your own trust boundary. Provided under GPL-3.0 without warranty; see LICENSE.

Project status

Deployed and operating on the maintainer's infrastructure, including a dual-surface content-transfer model (agent surface carries no content tools; file content moves only through a separate token-gated CLI surface). The wall, grant store, and audit chain are production-tested; the development discipline is adversarial review with runnable attack scripts in review/.