Point recon at a domain and get its public technology and identity footprint: email security, mail and identity providers, SaaS indicators, and certificate-transparency findings. No credentials, no API keys, no active scanning. Ships as a CLI, versioned JSON, and a local MCP server for agent tools. Lookup time depends on public-provider latency and whether certificate transparency enrichment is enabled.
The panel above is generated, not captured:
scripts/generate_terminal_demo.py
drives recon's real formatter over a deterministic, no-network fixture for the
fictional Example Industries Ltd. It shows the shape of a full-signal
result, every row a rich target can fill, and no live lookup of reserved
example.com reproduces it. No real organization is depicted. Other project
fixtures use IETF reserved .invalid namespaces.
Accessible text transcript
demo@recon:~$ recon example.com # synthetic fixture, not a captured run
Example Industries Ltd
example.com
──────────────────────────────────────────────────────────────────────────────
Provider Microsoft 365 + Proofpoint gateway
Tenant a1b2c3d4-e5f6-7890-abcd-ef1234567890 • NA
Tenant domain example-industries.onmicrosoft.example.com
Auth Federated
Confidence ●●● High (4 sources)
Services
Email Microsoft 365, Proofpoint, DMARC reject, DKIM,
SPF strict, MTA-STS enforce
Identity Okta
Cloud Cloudflare (CDN/edge), AWS Route 53 (DNS)
Security Wiz Security
Data & Analytics Snowflake, Datadog
Collaboration Slack, Atlassian (Jira/Confluence), GitHub, Zoom
Evidence roles: --explain
High-signal related domains
login.example.com, support.example.com, status.example.com
Insights
Federated identity observed; identity-vendor indicators: Okta
Email security: observed controls: DMARC reject, DKIM, SPF strict, MTA-STS
Defensive use only. Posture review, vendor diligence, architecture review. See docs/legal.md.
macOS / Linux:
curl -fsSL https://raw.githubusercontent.com/blisspixel/recon/main/scripts/install.sh | bashWindows (PowerShell):
powershell -NoProfile -ExecutionPolicy Bypass -Command "irm https://raw.githubusercontent.com/blisspixel/recon/main/scripts/install.ps1 | iex"The installer sets up recon for your user account, including uv and Python
3.14 when needed. No administrator access or existing Python setup is required
for a fresh install. Open a new terminal afterward.
Already have Python 3.11 through 3.14? These work too:
pip install recon-tool
# Or install in an isolated tool environment:
uv tool install recon-tool
pipx install recon-toolIf your system Python is externally managed, use the installer above or a virtual environment.
recon --version # offline install check
recon doctor # check installation, latest release, and online connectivity
recon update # update through the original package manager
recon update --check # check for a newer release without installingOn Windows, recon update starts a background update after the running launcher
exits. It prints the path to a local progress and result log. Once complete,
run recon --version to confirm.
The scripts install an exact published version, preserve a sole existing uv
or pipx owner, and explain how to recover ambiguous or unmanaged installs.
For a fixed installer, download a
release tag source archive
and review scripts/install.sh or scripts/install.ps1 before running it.
Artifact verification is covered in the
consumer verification recipe.
Before the first lookup, know what leaves your machine. recon makes DNS queries
that recursive and authoritative DNS infrastructure may observe. Its only
default request to a target-owned endpoint is the standards-defined MTA-STS
policy fetch; Google CSE and BIMI certificate probes run only when
--direct-probes is explicitly enabled. See
ADR-0011.
Run a real review by replacing the placeholder with the domain you want to review:
recon "<domain-you-want-to-review>"For a real lookup showing the output shape, you can use recon example.com. recon
ships no offline demo mode: this is a real lookup command, but an ordinary
lookup may reuse a recent result from the default 24-hour cache. Pass
--no-cache for fresh collection. Reserved names such as example.com return a
panel of stray public residue from unrelated test configurations, including a
meaningless display name, at High confidence. It shows the shape of the output,
not a result about any organization.
A domain is a query coordinate, not proof of one organization or product: recon reports observations, not verdicts. That is the caution to keep beside every row the panel shows.
Install, update, uninstall, and first-run detail: docs/getting-started.md.
| You are trying to | Start here | What you get |
|---|---|---|
| Review one namespace | recon "<domain>" |
A compact, hedged panel of observed mail, identity, service, and confidence signals |
| Create a deterministic evidence handoff | recon review "<domain>" |
One fresh, role-neutral NamespaceReviewBundle with collection validity, evidence-linked candidates, and a standing scope statement |
| Explain or hand off a finding | recon "<domain>" --explain |
Per-source status, lineage-qualified rules, and retained-evidence counts; use --json --explain for the structured evidence graph |
| Review an operator-supplied set | recon batch domains.txt --json --include-ecosystem --summary --summary-schema 2.2 |
One portfolio evidence bundle with ordered typed results and errors, observable ecosystem overlaps, and an aggregate cohort summary |
| Work through an agent | recon mcp install --client=<name> |
Local typed tools, resources, and prompts without a hosted recon service |
Defenders can follow the evidence-first review workflow. Consultants and analysts should treat every multi-domain set as operator supplied: recon describes observable similarities and differences but does not infer ownership, control, or a corporate relationship.
NamespaceReviewBundle v1
provides a role-neutral handoff for one namespace: one fresh,
lookup-result-cache-bypassed baseline, evidence-linked review candidates,
explicit bounded collection and failure states, and one deterministic human
rendering. Use recon review "<domain>" from the CLI or
build_review_bundle(domain) through MCP. Both surfaces return the same v1
artifact; the CLI renders it as Markdown unless JSON is requested.
| Need | Use recon for | Use something else when |
|---|---|---|
| Fast external stack context | Passive DNS, identity-endpoint, CT, SaaS, and posture indicators | You need authenticated tenant inventory or asset-management truth |
| Defensive review or vendor diligence | Hedged observations and evidence traces you can verify | You need vulnerability scanning, exploit checks, or host-level facts |
| Automation-friendly output | Stable JSON, batch mode, delta mode, and local MCP tools | You need dashboards, scheduling, or report generation built in |
recon reports observations, not verdicts. Public channel ceiling: docs/limitations.md. How to report a result without overstating it: docs/reporting-observations.md.
recon is designed around a harder question than record collection: when may several incomplete public observations support one claim, and when must the result stay unresolved? The correlation model is the deeper design contract for that question. It separates observed facts, deterministic deductions, and model-relative diagnostics; treats provenance and abstention as first-class; and keeps ownership, causation, complete-inventory, and unvalidated robustness claims outside the result.
The numerical diagnostics have a narrower meaning than their precision may
suggest. slug_confidences measures evidence strength; a Bayesian posterior
such as 0.9314 is conditional on manually encoded model assumptions, not a
validated 93% chance of product use. Its uncertainty band is a display heuristic,
not a confidence or credible interval. No current claim family has the
independent validation needed for general calibration claims; see
statistical assurance.
The general robustness machinery in correlation section 5 remains research.
v2.15 and v2.16 closed a five-round presentation-drift class. Independent testers installed the published package, never read the source, and kept finding the same issue: a decision applied to one renderer and not the others. The current engine surface and core contracts are stable. Those findings and their fixes are in CHANGELOG.md and in ADR-0015 through ADR-0017.
The living work is the fingerprint catalog. Vendors add, rename, and retire the
public patterns recon detects, so a rule with no re-check is a slow source of
false positives and negatives. Useful contributions are scoped catalog changes
with a current exact vendor reference (or an independent disclosure-safe basis),
a genuine verified review date, synthetic positive/lookalike/sparse fixtures,
and provenance tests. A vendor-looking string alone is not sufficient support.
CONTRIBUTING.md
has the schema, the validation command, and what is deliberately out of scope.
The freshness loop and its coverage floor live in
docs/catalog-strategy.md.
For an agent-assisted maintainer pass, use the
corpus planning and catalog-round workflow.
It separates offline selection, explicitly bounded collection, candidate
refinement, and untouched holdout evaluation. Increased classification coverage
does not by itself establish accuracy or justify promotion.
The next improvement loop combines documented catalog coverage with automated checks for duplicate evidence, deceptive lookalikes, failed sources, and output agreement. Local implementation within an authorized task can proceed without human approval steps. Per-slug evidence strength now counts each exact source-type/value pair once, even when multiple rules match it; distinct pairs still use the existing heuristic. The acceptance loop sets the next work and its executable pass conditions.
recon example.com # default panel
recon example.com --explain # evidence trail
recon example.com --gaps # neutral hardening prompts
recon example.com --plain # panel as linear text (screen readers, grep)
recon example.com --plain --full # every field, linear
recon example.com --json # structured record
recon review example.com # deterministic namespace review
recon review example.com --output review.json # save the validated v1 artifact
recon example.com --explain-dag --explain-dag-format mermaid # Bayesian model DAG
recon batch domains.txt --json # batch JSON array
recon batch domains.txt --json --include-ecosystem --summary --summary-schema 2.2
# portfolio evidence bundle
recon delta example.com # diff vs local cache
recon capsule capture example.com -o run.json # caller-owned replay artifact
recon mcp install --client=cursor # wire MCP into a clientMore flags: docs/cli-surface.md. JSON contracts: schema · review bundles · stability · operational contract.
Versioned JSON remains recon's structured runtime contract. The Open Knowledge Format v0.2 projection is deferred until a named consumer justifies the mapping. recon does not emit OKF, and any future OKF view would be additive rather than a replacement for JSON. Caller-owned JSON observation capsules are documented in docs/observation-capsules.md, with the decision boundary in ADR-0014.
docs/surface-inventory.json, docs/cli-surface.md, and
recon://surface-inventory are generated discovery context and drift guards,
not stable runtime API contracts. ADR-0007 records the promotion gate.
Wire recon into Claude Desktop, Claude Code, Cursor, VS Code, Windsurf, Kiro, or any MCP-compatible client:
recon mcp install --client=claude-desktop
# also: claude-code, cursor, vscode, windsurf, kiro
recon mcp doctorRestart the client after installation, then run recon doctor --mcp,
recon mcp doctor, and recon doctor --client=<name> in that order. Start with
manual tool approvals. Treat agents as untrusted input.
Full setup:
docs/mcp.md.
Per-client scaffolds and skills:
agents/.
The PyPI package installs the CLI and MCP runtime. Repository agent scaffolds
and plugin files are not included in the wheel or release assets.
The current scaffolds use each client's native configuration. In particular,
the source-checkout-only Claude Code plugin uses Claude Code's client-specific
layout; it does not claim conformance with the portable
Agent Plugins v1.0.0 specification.
A complete-surface portable candidate now lives under
agents/agent-plugin/
and passes network-free validation against the exact pinned v1.0.0 schemas,
whose bytes still match the Published specification. That is not a client
compatibility or conformance claim. Representative-client evaluation remains
incomplete, so use recon mcp install --client=<name> as the released setup
path. The aggregate evaluation status and stop conditions are
recorded here.
| You say | What the agent should do |
|---|---|
| "Recon example.com" | Call lookup_tenant (or recon example.com) and return the panel-style summary |
| "What does example.com run for email and identity?" | Same lookup; lead with MX/IdP/tenant facts and confidence |
| "Why do you think that?" | Re-run with explain / provenance (--explain or explain=true) |
| "Compare example.com and example.net" | compare_postures or two lookups side by side |
| "Any public hardening gaps?" | find_hardening_gaps after a lookup - hedged "Consider" notes only |
Example chat. This transcript uses the same synthetic Example Industries
fixture as the illustration above, so it shows the shape of a full-signal
answer rather than what a live example.com lookup returns:
You: Recon example.com and tell me what you observe.
Agent: Public evidence for
example.com(fictional demo):
- Mail: Microsoft 365 delivery path plus Proofpoint gateway; DMARC
p=reject, SPF strict, MTA-STS enforce.- Identity: Federated; Okta indicator via CNAME.
- Tenant: GUID present, region NA; confidence high from four sources.
- Also seen: Cloudflare and Route 53 infrastructure, Wiz, Snowflake, Datadog, Slack, Atlassian, GitHub, and Zoom public indicators.
These are public-record observations, not a security grade or proof of active product use. Say if you want evidence detail or a gap review.
Guidance: agents/claude-code/skills/recon/, AGENTS.md.
Local execution is the default, and the project does not operate a hosted service. For teams that want shared remote access, the repository includes a draft authenticated container and Cloud Run Terraform starting point. It is intended to be directionally useful, not a validated production deployment. Operators own deployment, identity, data handling, cost, and operations.
| Topic | Link |
|---|---|
| Glossary of recon's terms | docs/glossary.md |
| Install and first commands | docs/getting-started.md |
| Reporting a result without overstating it | docs/reporting-observations.md |
| How it works | docs/how-it-works.md |
| Known weak areas and conservative wording | docs/weak-areas.md |
| Catalog growth and freshness loop | docs/catalog-strategy.md |
| Observation capsules | docs/observation-capsules.md |
| Correlation model | docs/correlation.md |
| MCP and agents | docs/mcp.md, agents/ |
| Full docs index | docs/README.md |
| Roadmap | ROADMAP.md · docs/roadmap.md · docs/strategic-gap-audit.md |
| Changelog | CHANGELOG.md |
| Security | SECURITY.md · docs/security.md |
Research and publication pointers (maintainer track, not the product core): docs/submission-freeze-checklist.md, validation/2026-06-30-submission-freeze-local-proof.md, docs/public-label-snapshot-decision.md (public lists as robustness checks rather than population rates), and docs/m365-tenancy-decision.md.
Repository and release tasks use uv 0.11.17. The broader
>=0.11.8,<0.13 project constraint exists for dependency-update tooling, not
as the reproducible maintainer toolchain.
uv self update 0.11.17
uv --version # must report uv 0.11.17
uv sync
uv run pre-commit install
uv run pre-commit install --hook-type commit-msg # required; see below
uv run python scripts/release_readiness.py --allow-dirty
uv run python scripts/check.pyThe commit-msg hook is a separate installation step and is not optional.
Attribution trailers live in commit metadata rather than in a diff, so no
file-content hook can see them, and a trailer that reaches a pushed branch
cannot be removed afterward: GitHub retains the head commit of every pull
request under refs/pull/*/head, where the repository owner cannot delete it.
AGENTS.md
states the rule.
uv run python scripts/check.py is the canonical local gate: lint, type
checks, coverage-gated tests, generated-artifact and catalog checks, text and
link hygiene, interface and claim checks, and size and complexity ratchets.
It mirrors the blocking core CI gate; hosted scheduled and supply-chain
workflows remain separate checks.
Project hygiene: keep examples reserved and synthetic, keep validation artifacts aggregate-only, and avoid dead code or placeholders. See CONTRIBUTING.md.
Apache 2.0. Free to use, build on, fork, and share. See LICENSE for the full terms.