MCP · Model Context Protocol21 read-only tools

Give an agent the whole evidence base

The Second Chair corpus is available to AI agents through a live Model Context Protocol server. Add it as a connector and your agent can search 43,432 passages — every open-distribution source; the 124 restricted works are analyzed but their text is never returned — walk the topic map, split any cluster by voice or era, find the rare places where voices actually disagree, and audit each claim against the evidence its author offered for it. We are applying analytical lenses to the corpus: topic mapping and actor mapping are published, with driver and dynamics mapping and scenario mapping in progress.

Connector URLhttps://insightbridge-secondchair.aicolab.org/mcp

Claude — personal accounts (Free, Pro and Max)

Go to Customize → Connectors — or open claude.ai/customize/connectors directly.
Click +, then Add custom connector.
Paste the URL above and click Add. No authentication is required — leave Advanced settings empty.
In a chat, click + (or type /), hover Connectors, and toggle it on for that conversation if it isn’t already.

Connectors are account-level: added once, available in Claude on the web, desktop and mobile. Free accounts can add one custom connector.

Claude — organisation accounts (Team and Enterprise)

On Team and Enterprise plans only an organisation Owner (or Primary Owner) can add a custom connector — members do not see the option, so the personal steps above will dead-end.

An Owner opens Organization settings → Connectors — or claude.ai/admin-settings/connectors directly.
Click Add, hover Custom, then select Web.
Paste the URL above and click Add. No authentication is required — leave Advanced settings empty.
Each member then opens Customize → Connectors, finds the connector the Owner added, and clicks Connect; in a chat it is enabled via + → Connectors like any other tool.

ChatGPT — requires developer mode

Turn on Developer mode. OpenAI has moved this toggle more than once: look under Settings → Apps → Advanced settings, and if it is not there try Settings → Connectors → Advanced or Settings → Security and login. On Business, Enterprise and Edu workspaces an admin has to allow it under Permissions & Roles → Connected data.
Go to Settings → Connectors. The button to create a custom connector only appears once developer mode is on.
Add the URL above with transport Streamable HTTP and authentication None.

Available on Plus and Pro; on workspace plans an administrator can disable developer mode or allowlist specific connectors.

Claude Code

One command — --transport http is what makes it a remote connector rather than a local process:

claude mcp add --transport http second-chair https://insightbridge-secondchair.aicolab.org/mcp

Re-adding after a change needs claude mcp remove second-chair first.

VS Code (GitHub Copilot), Cursor and Windsurf

All three take the same shape of entry. In VS Code run MCP: Open User Configuration from the command palette, or create .vscode/mcp.json in a workspace; in Cursor use Settings → MCP.

{
  "servers": {
    "second-chair": {
      "type": "http",
      "url": "https://insightbridge-secondchair.aicolab.org/mcp"
    }
  }
}

Some clients name the top-level key mcpServers rather than servers; if the server does not appear, try the other spelling.

Then ask something like “Using the legal-AI corpus, which claims about AI replacing lawyers are made by vendors, and what evidence do they actually offer for them?”

Works with any MCP client. The endpoint speaks JSON-RPC over a stateless one-POST Streamable-HTTP adapter at POST /mcp; there is no SSE stream and no authentication, so a client that insists on OAuth should be set to “no auth”.

Reading rules

How to read what comes back

Source counts are not prevalence. This is a curated evidence base of 1,501 published works, not a survey of the profession. "N sources say X" measures the corpus, never the field.
Positions are relative to each cluster's own proposition — Supports, Builds on, Unclear, Mixed, Redirects, Opposes measure stance-toward-a-framing. Most read supportive, so agreement is the default and only contested_topics finds real divergence.
Era is ordinal, not a category: Foundations (pre-2015) → Automation debate (2015-2022) → Generative shock (2023-2024) → Adoption at scale (2025) → Current wave (2026). The tools always return era rows in that order.
The tree has generations. list_families is the root generation, list_superclusters the one below; any node's id works with topics_in_supercluster. Every topic carries its lineage.
Restricted sources are analysed, not reproduced. Their stances, key points and short verified quotes are returned; search_passages never returns their text.
topic_id is the topicClusterId, not the display cluster_id; both look plausible, so the wrong one quietly returns a different topic.

Live tool tester

Try the tools on Second Chair

Choose a tool, adjust its example inputs, and run it against the live server on Second Chair.

Open Second Chair’s live tool tester (external site) ↗

The tools

What an agent can ask for

Twenty-one read-only tools. Each carries the corpus’s reading rules in its own description.

Orientation & search

corpus_overview

Scale and shape of the corpus: 1,501 sources, the topic tree, the two lenses, the position mix. Read first.

search_passages

Semantic search over the verbatim passages of open-distribution sources; restricted works are never returned.

find_topics

Semantic search over the 369 topic clusters; returns the ids the tools below need, with their lineage.

The grouping tree

list_families

The 5 top-level families — the coarsest map of the corpus — with child, topic and source counts.

list_superclusters

The 46 named themes one generation below the families, each with its parent family and reach.

topics_in_supercluster

Every topic beneath a family or theme id, each with its source reach, position mix and lineage.

related_topics

Topics sharing a nearest-generation group with the given topic, ranked by how many groups they share.

Anatomy of a topic

get_topic

One cluster in full: proposition, findings, lineage, reach by tier, position mix by voice and era, stances.

topic_quotes

Verbatim cross-cutting quotes for a topic, each attributed to its source, voice, era and publication.

topic_perspectives

The parameterized lens: each voice’s synthesized stance, or each era’s — era rows in ordinal order.

Divergence & drift

contested_topics

The corpus’s disagreements: clusters ranked by how many stances push back, each naming its dissenters.

era_drift

How positions move across the five eras — for one topic, one voice, both, or the whole corpus.

Voices & sources

source_perspectives

Every source holding a stance on a topic, with its framing and own points — slice by voice, era, position.

voice_profile

One voice across the corpus: eras, evidence types, position mix, and the clusters it dominates.

list_sources

The 1,501 published works — filter by voice, era, class, type, year or a substring of title or slug.

get_source

One work by slug or fuzzy title: metadata, every position it holds, its facets, its version siblings.

Documents

find_documents

Search the 1,501 documents by title, source slug, publication or author. Documents are 1:1 with sources.

get_document

Key points with verified quotes, the full claim spine, the three curator facets, dominant topic.

Tags & the claim audit

list_tags

One curator vocabulary — evidence type, topic dimension, jurisdiction — with source counts per value.

tag_pivot

Source counts per tag value, cross-tabbed by voice, era, type or class; evidence type × voice is sharpest.

claim_audit

Each concrete claim paired with the evidence offered, its timeframe and who it affects; filter to unevidenced.

How this was made

Second Chair is a jointly developed corpus. The corpus underlying this project was created by Anthea Roberts and David B. Wilkins. It was structured and analyzed using Insight Bridge, the corpus-reading pipeline built by Sam Bide at AI CoLab, which reads each source end-to-end, embeds its passages and clusters them into topics, themes and families. It was then analyzed through Dragonfly Thinking’s multi-perspective agentic method. How the reading was produced →