Deterministic B2B context router. Zero keys. Schema-locked JSON your agent pipelines can actually trust.
Project description
companyctx
Deterministic B2B context router. Zero keys. Schema-locked JSON your agent pipelines can actually trust.
pipx install companyctx # v0.2.0 on PyPI — schema_version field + structured errors
companyctx fetch acme-bakery.com --json
{
"data": {
"fetched_at": "2026-04-22T18:35:02.767112Z",
"mentions": null,
"pages": {
"about_text": "Acme Bakery has served Portland, OR since 2010. ...",
"homepage_text": "Acme Bakery is a bakery in Portland, OR. Founded 2010. We're a team of 3. ...",
"services": ["Custom cakes", "Catering", "Wholesale bread", "Pastry boxes"],
"tech_stack": ["WordPress", "Elementor"]
},
"reviews": null,
"signals": null,
"site": "acme-bakery.com",
"social": null
},
"error": null,
"provenance": {
"site_text_trafilatura": {
"cost_incurred": 0,
"error": null,
"latency_ms": 412,
"provider_version": "0.1.0",
"status": "ok"
}
},
"schema_version": "0.3.0",
"status": "ok"
}
One site in. One schema-locked JSON object out. No API keys for the
zero-key path. Graceful partials on anti-bot blocks. The envelope is
versioned via a top-level schema_version field so agents can branch on
shape without substring-parsing.
reviews / social / signals / mentions are reserved in the
CompanyContext schema; the providers that populate them (Google Places,
YouTube Data, the site-heuristic provider) are roadmap work — see
Status below. Today a real run returns pages.* plus nulls,
which is what the --mock fixture tree reproduces byte-for-byte.
Status
v0.2.0 is the current release. What's shipped:
- Envelope contract —
{schema_version, status, data, provenance, error?}withstatus ∈ {ok, partial, degraded}and a structuredEnvelopeError({code, message, suggestion}) on non-okruns.extra="forbid"on every model. Seedocs/SCHEMA.md. - Zero-key Attempt 1 —
site_text_trafilatura(TLS-impersonated fetch viacurl_cffi+ trafilatura extraction). Populatesdata.pages.*(homepage_text,about_text,services,tech_stack). 20/20 envelope-okon a shape-balanced zero-key probe at a fresh Chrome fingerprint; seedocs/ZERO-KEY.md. - Smart-proxy Attempt 2 —
smart_proxy_http(vendor-agnostic URL-style proxy). User suppliesCOMPANYCTX_SMART_PROXY_URL; env-unset surfaces asnot_configuredwith an actionable suggestion rather than a crash. Named- vendor adapter lands after the smart-proxy vendor eval (tracking: #63). - CLI verbs —
fetch,schema(emits Draft 2020-12 JSON Schema),validate,providers list(with--json, waterfall tier, config status, reason). py.typedmarker + public-API re-exports —from companyctx import Envelope, EnvelopeError, CompanyContext, ...works straight off the package root; downstreammypysees concrete types.
What's not shipped yet and is explicitly deferred (the CLI surface rejects these flags with a tracking-issue pointer rather than silently accepting them):
- SQLite fetch cache (
--from-cache/--refresh/--no-cache/--config) — tracked in #9. batch <csv>— stub, prints an error and exits non-zero.- Direct-API (Attempt 3) providers — Google Places (tracking: #7),
Yelp Fusion, YouTube Data, Brave Search (mentions tracking: #58). None
registered in v0.2;
data.reviews/data.social/data.mentionsstay null on live runs until a direct-API provider lands. - Site-heuristic
signalsprovider — planned alongside the direct-API slate; populatesdata.signals.copyright_year,last_blog_post_at,team_size_claim. companyctx fetch --markdown— experimental flag labelled in help text; runs fail fast. Markdown output belongs in a downstream synthesis step, not here (tracking: #68).
Measured 97% envelope-ok rate on a 100-site real-world sample on the
v0.1 zero-key baseline; post-v0.2 re-measurement tracks under the durability
rerun (#61). Full breakdown in
research/2026-04-21-tls-impersonation-spike.md
and docs/ZERO-KEY.md.
Schema + CLI surface in docs/SPEC.md and
docs/SCHEMA.md; architecture in
docs/ARCHITECTURE.md;
release-readiness ADR in
decisions/2026-04-21-v0.1.0-release-readiness.md.
What this is (and isn't)
IS
- A schema-locked context router. The Pydantic v2 envelope is the product. Providers come and go; the contract agents consume does not.
- A Deterministic Waterfall. Zero-key stealth fetch first → smart-proxy provider (if configured) → direct-API provider (if configured). Every attempt returns the same shape.
- A narrow muscle in the brains-and-muscles pattern. Your frontier
model is the brain;
companyctxis one of many CLIs it pipes through. - A local-first memory layer — by design, landing later. The SQLite
cache is scoped as Vertical Memory: a queryable local B2B dataset that
compounds as a byproduct of normal use. It is not wired in v0.2 —
--from-cache/--refreshare deferred to #9.
ISN'T
- Not a scraper competing on scale. Residential-proxy / headless-browser
infrastructure is a commodity layer — we compose it via a
SmartProxyProviderinterface, we don't out-build it. - Not an agent framework. Orchestration lives upstream.
- Not a hosted service. Local pipx CLI. No rented infra. No credits.
- Not a synthesis engine. Our output is the input for synthesis.
- Not a Cloudflare bypass. Zero-key covers the majority of small-biz homepages. It won't defeat serious anti-bot — see the coverage matrix.
- Not a people-data tool. Companies only. Contact enrichment belongs upstream (Apollo, Clearbit, manual).
- Not an MCP server — ever, in our roadmap. MCP's ~50k-token schema
dump defeats a muscle built to save tokens. Agents find us via
SKILL.md; the CLI +jq+ stdout are the composition layer. Seedecisions/2026-04-20-skill-md-not-mcp.md.
The Deterministic Waterfall
Attempt 1 — Zero-key stealth (TLS+HTTP/2 impersonation + trafilatura + extruct)
↓ 403 / challenge / timeout?
Attempt 2 — Smart-proxy provider (user-keyed, vendor-agnostic)
↓ still blocked?
Attempt 3 — Direct-API provider (user-keyed — Google Places, Yelp Fusion, YouTube)
↓
{ schema_version, status, data, provenance, error? }
Every attempt maps to the same Pydantic schema. Downstream pipelines
never branch on which attempt succeeded — they branch on the envelope's
status: ok | partial | degraded.
On full block with no Attempt-2/3 providers configured:
{
"data": { "fetched_at": "...", "mentions": null, "pages": null, "reviews": null, "signals": null, "site": "example.com", "social": null },
"error": {
"code": "blocked_by_antibot",
"message": "blocked_by_antibot (HTTP 403)",
"suggestion": "configure a smart-proxy provider key or skip this prospect"
},
"provenance": { "site_text_trafilatura": { "cost_incurred": 0, "error": "blocked_by_antibot (HTTP 403)", "latency_ms": 842, "provider_version": "0.1.0", "status": "failed" } },
"schema_version": "0.3.0",
"status": "partial"
}
Never raises. Never crashes your pipeline. Every run comes back well-formed.
See docs/ARCHITECTURE.md for the full picture and
docs/ZERO-KEY.md for honest anti-bot scoping.
Zero-key coverage — honestly
| Site class | Zero-key outcome |
|---|---|
| Small-biz WordPress / Squarespace / Wix / Webflow / agency custom | Full payload. Measured 20/20 status: "ok" on a 20-site probe across eight ICP niches at the latest Chrome fingerprint.* |
| Cloudflare Turnstile / DataDome / Akamai / PerimeterX | Clear at a fresh fingerprint; flips to blocked_by_antibot (HTTP 403) as the pinned fingerprint ages. Returns status: "partial" with actionable suggestion. |
| JS-heavy SPAs needing a real browser | HTML shell only. Render-dependent fields come back null. Configure a smart-proxy provider to fill the gap. |
| Aggregator pages (Yelp / Houzz / G2 / Birdeye) | Not the target — use the direct-API providers (reviews_google_places, reviews_yelp_fusion) instead. |
* Stealth fetcher: curl_cffi pinned to impersonate="chrome146". Spike
method, raw JSONL, and stale-fingerprint decay analysis in
docs/ZERO-KEY.md +
research/2026-04-21-tls-impersonation-spike.md.
Numbers come from measurement, not marketing.
Brains-and-muscles pipe
companyctx fetch acme-bakery.com --json \
| jq '.data.pages | {services, tech_stack, homepage: .homepage_text[:280]}' \
| claude -p "write a 6-section outreach brief from this context"
import json, subprocess
ctx = json.loads(subprocess.check_output(["companyctx", "fetch", "acme-bakery.com", "--json"]))
if ctx["status"] != "ok":
err = ctx["error"] or {}
print(f"heads up: {err.get('code')} — {err.get('message')} → {err.get('suggestion')}")
brief = synthesize(ctx["data"]) # your synthesis call, your prompts, your weights
Branch on status (not on try/except). Branch on error.code (closed
enum), not on substring-matching error.message.
companyctx never calls an LLM. The brain upstream decides what the
context means.
Install
v0.2.0 is the current release (schema-breaking envelope bump):
pipx install companyctx
companyctx --version # companyctx 0.2.0
companyctx fetch acme-bakery.com --mock --json
From source:
git clone https://github.com/dmthepm/companyctx.git
cd companyctx
pip install -e ".[dev,extract,reviews,youtube]"
companyctx --help
Design invariants
- Schema is the product. Providers are replaceable; the
CompanyContextenvelope is not. Raw observations only — inference lives in the downstream synthesis layer. - Graceful-partial always. Providers never raise uncaught. Every
failure maps to
ProviderRunMetadata.statusper provider and the top-levelstatuson the envelope. - Vertical Memory (design invariant, not yet wired). The SQLite cache
is designed to compound into a queryable local B2B dataset under
XDG paths; the
schema-versioned migrations and
--refresh/--from-cacheflags are reserved in the CLI surface but not implemented in v0.2 — they reject with a tracking-issue pointer (#9). - Provider pluggability. Every deterministic call class is discovered
via Python entry points (
companyctx.providers). v0.2 shipssite_text_trafilatura(zero-key) andsmart_proxy_http(user-keyed, vendor-agnostic). Direct-API providers (Google Places, Yelp, YouTube, Brave) and thereadability-lxmlbus-factor fallback are scaffolded for later milestones. Seedocs/PROVIDERS.md. - robots.txt respected by default.
--ignore-robotsis an explicit CLI-only flag; never settable via TOML or env. - Deterministic mocks.
fixtures/<site>/drives--mock; re-runs produce byte-identical output modulofetched_at.
Providers
Shipped in v0.2 (run companyctx providers list to introspect):
| Slug | Tier | Category | Key | Cost |
|---|---|---|---|---|
site_text_trafilatura |
zero-key | site_text | — | free |
smart_proxy_http |
smart-proxy | smart_proxy | COMPANYCTX_SMART_PROXY_URL |
per-call |
Scaffolded / deferred (roadmap):
| Slug | Tier | Category | Tracking |
|---|---|---|---|
site_text_readability |
zero-key | site_text (fallback) | bus-factor fallback, lands alongside fallback wiring |
site_meta_extruct |
zero-key | site_meta | future milestone |
social_discovery_site |
zero-key | social_discovery | future milestone |
signals_site_heuristic |
zero-key | signals | future milestone |
reviews_google_places |
direct-api | reviews | #7 |
reviews_yelp_fusion |
direct-api | reviews | future milestone |
social_counts_youtube |
direct-api | social_counts | future milestone |
mentions_brave_stub |
direct-api | mentions | #58 |
Full rationale + the SmartProxyProvider interface in
docs/PROVIDERS.md.
Layout
companyctx/ # package
cli.py # Typer app
core.py # Deterministic Waterfall orchestrator (Attempt 1 + 2)
schema.py # pydantic v2 models — the JSON contract
extract.py # HTML → SiteSignals extractor shared by zero-key + smart-proxy
http.py # stealth fetcher helpers
robots.py # robots.txt enforcement
security.py # SSRF + response-cap + redirect guardrails
providers/
__init__.py # plugin loader (importlib.metadata.entry_points)
base.py # ProviderBase protocol + FetchContext + Literal category enum
site_text_trafilatura.py # Attempt 1 — zero-key homepage extraction
smart_proxy_http.py # Attempt 2 — vendor-agnostic smart proxy
smart_proxy_base.py # shared helpers for smart-proxy providers
SKILL.md # ~150-token agent-discovery surface (not MCP)
docs/
SPEC.md # frozen spec snapshot (v0.2 envelope)
SCHEMA.md # Pydantic envelope in detail
ARCHITECTURE.md # brains-and-muscles + Deterministic Waterfall + Vertical Memory
ZERO-KEY.md # honest anti-bot coverage + graceful-partial contract
PROVIDERS.md # provider list + SmartProxyProvider interface
VALIDATION.md # two-phase acceptance protocol
REFERENCES.md # upstream OSS deps
decisions/ # in-repo ADRs (walks-the-walk for OSS readers)
fixtures/ # per-site raw HTML + expected.json (drives --mock)
tests/ # pytest, hypothesis where useful
Contributing
See CONTRIBUTING.md. Conventional commits, ~400 LOC per
PR, one PR per milestone or provider. ruff + mypy strict + pytest ≥70% cov.
Architecture-shape changes go through decisions/ first.
License
MIT. Copyright 2026 Noontide Collective LLC.
Support
companyctx is open source and free to use. If it earns a place in your
pipeline, consider supporting development by joining Noontide's Main Branch
community on Skool (there's a free trial): https://skool.com/main.
Project details
Release history Release notifications | RSS feed
Download files
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.
Source Distribution
Built Distribution
Filter files by name, interpreter, ABI, and platform.
If you're not sure about the file name format, learn more about wheel file names.
Copy a direct link to the current filters
File details
Details for the file companyctx-0.3.0.tar.gz.
File metadata
- Download URL: companyctx-0.3.0.tar.gz
- Upload date:
- Size: 86.5 kB
- Tags: Source
- Uploaded using Trusted Publishing? Yes
- Uploaded via: twine/6.1.0 CPython/3.13.12
File hashes
| Algorithm | Hash digest | |
|---|---|---|
| SHA256 |
38448e7775845113c5f00c07a086f29b66db89f99a50babbcd5c0ce34cd7e809
|
|
| MD5 |
4c005d8df1231503eaa6307cd3f04f5d
|
|
| BLAKE2b-256 |
5f0a88ece7da8bf471f1af93ceff8942832ffa9978d6c7a63ee7f01eff157db2
|
Provenance
The following attestation bundles were made for companyctx-0.3.0.tar.gz:
Publisher:
publish.yml on dmthepm/companyctx
-
Statement:
-
Statement type:
https://in-toto.io/Statement/v1 -
Predicate type:
https://docs.pypi.org/attestations/publish/v1 -
Subject name:
companyctx-0.3.0.tar.gz -
Subject digest:
38448e7775845113c5f00c07a086f29b66db89f99a50babbcd5c0ce34cd7e809 - Sigstore transparency entry: 1361070943
- Sigstore integration time:
-
Permalink:
dmthepm/companyctx@2dd9ead06d179b3320bed90d66345b257119a5e4 -
Branch / Tag:
refs/tags/v0.3.0 - Owner: https://github.com/dmthepm
-
Access:
public
-
Token Issuer:
https://token.actions.githubusercontent.com -
Runner Environment:
github-hosted -
Publication workflow:
publish.yml@2dd9ead06d179b3320bed90d66345b257119a5e4 -
Trigger Event:
release
-
Statement type:
File details
Details for the file companyctx-0.3.0-py3-none-any.whl.
File metadata
- Download URL: companyctx-0.3.0-py3-none-any.whl
- Upload date:
- Size: 52.3 kB
- Tags: Python 3
- Uploaded using Trusted Publishing? Yes
- Uploaded via: twine/6.1.0 CPython/3.13.12
File hashes
| Algorithm | Hash digest | |
|---|---|---|
| SHA256 |
cf3c6f004b628060462e914321077dfd9c0be125a70a4c81ca8a48276fdce37c
|
|
| MD5 |
c0b1eba692cc07a5415046240f654e66
|
|
| BLAKE2b-256 |
74fc536e8c7d81ddf9982416417b6d6d6709266b55af3ced1fa03ca0f2bd2e63
|
Provenance
The following attestation bundles were made for companyctx-0.3.0-py3-none-any.whl:
Publisher:
publish.yml on dmthepm/companyctx
-
Statement:
-
Statement type:
https://in-toto.io/Statement/v1 -
Predicate type:
https://docs.pypi.org/attestations/publish/v1 -
Subject name:
companyctx-0.3.0-py3-none-any.whl -
Subject digest:
cf3c6f004b628060462e914321077dfd9c0be125a70a4c81ca8a48276fdce37c - Sigstore transparency entry: 1361070950
- Sigstore integration time:
-
Permalink:
dmthepm/companyctx@2dd9ead06d179b3320bed90d66345b257119a5e4 -
Branch / Tag:
refs/tags/v0.3.0 - Owner: https://github.com/dmthepm
-
Access:
public
-
Token Issuer:
https://token.actions.githubusercontent.com -
Runner Environment:
github-hosted -
Publication workflow:
publish.yml@2dd9ead06d179b3320bed90d66345b257119a5e4 -
Trigger Event:
release
-
Statement type: