One context engine. Multiple permitted ways in.
A proposal to assess a reusable context engine across browser-native WebMCP, remote MCP, extension capture, files/connectors and manual input.
No adapter, tool contract, retention policy, deployment or integration path is approved by this document. Final decisions must be recorded in GitHub.
Recommendation for discussion
Keep one reusable context engine and expose it through adapters. The engine could normalise permitted inputs, classify and reconcile content, preserve provenance and freshness, identify trust and sensitivity, create task-aware context packs, validate them, and emit processing receipts. The first release should select the smallest combination that satisfies one measurable use case rather than implement every adapter.
Integration options
| Path | Best fit | Trade-off / control |
|---|---|---|
| Browser-native WebMCP | Compatible websites and clients | Not universally available; never make it the only path. |
| Remote MCP service | Reusable agent-facing tools across compatible clients | Requires hosting, identity, permissions and data-handling decisions. |
| Browser extension / user-mediated capture | Permitted pages without WebMCP | Requires explicit browser permissions and compatibility testing. |
| Files, clipboard and SaaS connectors | Documents, exports and API-accessible systems | Source-specific authorisation and adapters. |
| Manual paste or upload | Broad baseline when automation is unavailable or inappropriate | More friction; source metadata may need user confirmation. |
Suggested first fallback: manual paste/upload with source URL and capture time where available. Assess extension or connector support only after the first use case and permission model are agreed.
Proposed logical tool surface
context.search — search permitted sources for task-relevant contextcontext.capture — capture and normalise a permitted sourcecontext.inspect — return provenance, evidence and classificationcontext.reconcile — link duplicates/variants and surface contradictionscontext.pack.create — build a task-specific context pack within a budgetcontext.pack.validate — check permissions, provenance, freshness, conflicts and budgetcontext.pack.get — retrieve an existing packcontext.receipt.get — retrieve outcome and evidence for a processing runThese are proposed names, not existing endpoints. Owners must agree schemas, authentication, transport, versioning, errors and permissions before implementation.
Open decisions for MCP and WebMCP owners
- What is the first end-to-end use case and measurable success threshold?
- Should the first release expose remote MCP, browser-native WebMCP, or both?
- Which no-WebMCP fallback is in scope first?
- What is processed locally versus remotely, and what may be retained?
- How are users and agents identified, authorised, scoped and audited?
- What retention and deletion policy applies to raw captures, derived context, packs and receipts?
- Which fidelity, security and quality tests must pass before a context pack can be sent?
Acceptance tests
- A supported client discovers tools and uses the published schemas.
- An agreed no-WebMCP path submits content and source metadata to the same engine.
- A context pack preserves required facts, caveats, relationships and source references.
- Duplicate/variant linking preserves meaningful changes and measures false merges.
- Untrusted page text cannot silently become an authorised instruction.
- Users can inspect what was retained, omitted, merged and why.
- A receipt records input references, engine/schema version, outcome and measurements.
- Mocked, estimated and observed results are clearly distinguished.
Working principle
Remove repetition, not meaning. Compress context, not evidence. Captured content is not automatically authorised instruction.
Related implementation proof: browser integration issue · durable queue and isolation issue. See also the architecture guide and release proof gates.