About
Built around one refusal.
Reconnaissance tooling is excellent and reconnaissance workflow is not. The tools are single commands; the afternoon goes to ordering them, reconciling nine output shapes into one picture, and writing up a result whose gaps only you know about.
Zero-Recon takes that orchestration on. The part it will not do is round a gap up into a clean result — which is the constraint most of the architecture below exists to enforce.
Principles
Four rules the code is held to.
These are not aspirations. Each one is a place where the obvious shortcut was available and was not taken.
- 01
Coverage before conclusions
How much of the plan completed is counted first, arithmetically, and every downstream statement is gated on it. Below full coverage the report says what it does not know rather than scoring what it never measured.
- 02
The gate lives below the UI
An active tool is refused by the executor, not by a disabled button. A control that can be clicked past is decoration, so the check sits underneath the API and underneath the interface.
- 03
Tools are data, not code paths
A tool is a registry row: an execution definition, input and output schemas, a cost model and a mandatory side-effect class. Adding one is a row, not a release.
- 04
No single model dependency
Language models plan and explain. They do not count coverage, decide authorization, or pick a target. The deterministic half keeps working when a provider does not.
Architecture
Separate processes, so the boundaries are real.
The isolation claims on this site are only worth anything if the pieces they describe are actually separate. They are separate services, not modules in one process trusting each other.
- Web
- The workspace, the dashboards and this site. Public pages render on the server and ship no application JavaScript.
- API
- Authentication, roles and permissions, tenancy, billing and the realtime hub. Every query is tenant-scoped at the repository layer.
- Orchestrator
- Builds the task graph, decides what is worth doing next from what has been observed, and holds the authorization token that the executor demands.
- Tool workers
- Each tool runs in its own short-lived container with a read-only root, dropped capabilities and hard resource limits. Arguments are templated into argv.
- Data
- The entity graph, missions, reports and the append-only audit log. Redis carries queues and realtime fan-out.
- Model providers
- Behind one interface, selected by configuration. Swapping providers is a setting, not a rewrite.
next.js
fastapi
planner · executor · replanner
one container per task
postgresql · pgvector · redis
gemini · openai · anthropic
What a tool container is given
- Filesystem
- read-only
- Capabilities
- all dropped
- Privilege escalation
- denied
- Network
- scoped
- Credentials
- none
- Shell
- never involved
Tool images are pinned by digest against an allowlist. An image that is not on it is refused rather than pulled.
Status
Where the platform actually is.
Zero-Recon is a preview build run by a small team. The honest version of a roadmap is three columns, and the middle one is the one most products leave out.
Shipping
In the product today
- Agent planning, approval and adaptive replanning
- Twenty-one tools, passive and active, in sandboxed containers
- Authorization records via DNS TXT domain verification
- One entity graph per target, deduplicated and confidence-scored
- Coverage-gated reports in Markdown and JSON
- Organizations with owner, admin, member and viewer roles
Deliberately not built
Choices, not gaps
- No arbitrary code execution and no tool the agent wrote itself
- No model chosen by the agent at runtime
- No score published above the coverage the mission actually reached
- No control shown enabled that the server will then refuse
Not yet
Honestly unfinished
- No self-serve checkout — Premium is arranged directly
- No third-party audit and no compliance certification
- Documentation is an overview, not an API reference
- Tool acquisition beyond the bundled set is curated and manual
Zero-Recon holds no compliance certification and has had no third-party security audit. The controls described across this site are ones you can read in the code and verify against your own instance — that is the whole of the claim.
Judge it against a target you own.
Passive reconnaissance runs on the free daily allowance. That is enough to see whether the report tells you the truth about itself.