Capabilities

Everything Documesh can do β€” as experienced by a human in the browser, and as callable by an AI agent through WebMCP. Every capability is a real tool with a schema, deterministic behavior, and full source attribution.


answer tool: explain_error

🩺 Errors-to-Docs Bridge

Paste any stack trace or error log. The bridge extracts error signatures β€” exception classes, k8s state reasons, errno codes, exit codes β€” and matches them against the index, returning the three closest official documentation sections with a fix sketch. Crucially, it declines to guess: every answer carries an explicit disclaimer that this is a closest-match, not a diagnosis, and unknown errors return labeled "closest matches" rather than confident nonsense.

Real example
Input:  Warning Failed kubelet Back-off restarting failed container…
        Error: CrashLoopBackOff: container failed to start

Output: signatures: ["CrashLoopBackOff", …]
matches:
  1. kubernetes β€” "How Pods handle problems with containers"
     β†’ Container restart policy   (score 252.4)
     https://kubernetes.io/docs/concepts/workloads/pods/pod-lifecycle/
  2. netlify   β€” "Fix a failed deploy"
Try it live β†’ | Evaluated: 5/5 curated real-world errors matched (100%)

capability built into every tool

🏷 Version awareness

Kubernetes docs are indexed per release branch (v1.32, v1.29) β€” so a question about pod lifecycle returns both versions side-by-side, letting you spot behavioral changes between releases. Every answer names its version. This is the anti-"2021 blog post" feature: the mesh grounds answers in the exact version you run.

Versioned answers in practice
search: "container restart policy"

kubernetes v1.32 β†’ "Container restart policy"   score 252.4
kubernetes v1.29 β†’ "Container restart policy"   score 236.4
                   ↑ same section, two releases β€” diffs become visible

capability structural, not promotional

βš–οΈ License-aware ingestion

Ingestion is gated on legality, not convenience. Vendors enter the mesh only via (a) permissive licenses (CC-BY-4.0, MIT, Apache-2.0) or (b) official agent interfaces (llms.txt, `.md` endpoints) that explicitly permit machine consumption. Vendors with unclear or restrictive terms β€” Terraform (BUSL), Grafana (AGPL) β€” are deliberately excluded. Every single response names its license; the coverage page documents both sides of that line.


powered by webmcp

πŸ€– Fully agent-operable

Everything above is exposed as WebMCP tools β€” structured, schema-described capabilities registered in the page via document.modelContext.registerTool(). A visitor's existing agent (ChatGPT in the in-app browser, or any MCP-compatible agent in Chrome) discovers and calls them directly. The human sees the same cited sources on screen β€” collaboration, not automation.

Full WebMCP tool reference β†’