## What this is for When a candidate enables MCP on a share link, you receive **two things** alongside the share URL: - The share URL itself — opens the candidate's Twin in a browser, lets you chat with it directly. - An **MCP token** that looks like `aigis_sk_`. This token plugs into your AI agent (Claude Desktop, Claude Code, or Claude.ai) so the agent can talk to the candidate's Twin on your behalf — calling tools like `aigis_score_against_jd`, `aigis_get_evidence_for`, and `aigis_get_narrative_anchors` directly. The MCP path is for when you want **your AI agent** to do the analysis, not when you want to read it yourself. Both paths work; pick whichever fits your workflow. ## What you can and can't do **Read-only by design.** The token grants access to the candidate's profile, knowledge base, and analysis tools. It cannot submit journal entries, modify the candidate's data, or take any action that would change what's stored on the candidate's side. If a recruiter workflow ever needs write access (e.g. submitting a hiring decision back to the candidate), it will require a structurally different consent flow — share-link MCP is intentionally locked to reads. **Per-link tool presets.** The candidate picks one of three presets when creating the link: - **Full read** (default) — every recruiter-facing read tool: chat (`ask_candidate`, `chat_with_twin`), JD analysis (`extract_jd_requirements`, `score_against_jd`), fact-check (`query_knowledge`, `get_evidence_for`), and profile/discovery. - **JD analysis** — fact-check primitives + JD scoring (`extract_jd_requirements`, `score_against_jd`) + profile/discovery. Excludes open chat. Best when you're scoring fit, not chatting freely. - **Fact-check only** — `get_evidence_for` + `query_knowledge` + profile/discovery. Tightest preset — for verifying specific claims rather than open-ended exploration. **Two tools are never exposed via recruiter share-links**, regardless of preset: - `aigis_prepare_for_role` — surfaces "likely questions the recruiter will ask" + the candidate's gaps. Built for the candidate to anticipate a recruiter's probing, not for the recruiter to read. - `aigis_get_narrative_anchors` — returns cover-letter feedstock (recurring themes + first-person hooks) for the candidate's own LLM to weave into a pitch. These are candidate-side tools — recruiters paying for fit-assessment shouldn't drive the candidate's narrative. They're available to a candidate's own AI agent via OAuth or API key (see [Connecting AI assistants](/help?tab=agents-mcp&article=connecting-claude)). If you call a tool that's not in the link's preset, the response is a structured `tool_not_in_preset` error with the list of allowed tools — your agent can fall back to a tool it's allowed to call. Available tools (all read-only): - `aigis_get_profile`, `aigis_get_certifications`, `aigis_get_availability`, `aigis_get_readiness`, `aigis_get_knowledge_summary`, `aigis_get_twin_capabilities`, `aigis_whoami` — basic candidate info / discovery (every preset) - `aigis_query_knowledge` — raw RAG chunks for your agent to synthesise (every preset) - `aigis_get_evidence_for` — fact-check a specific claim (every preset) - `aigis_extract_jd_requirements` + `aigis_score_against_jd` — JD analysis suite (`jd_analysis`, `full_read`) - `aigis_ask_candidate`, `aigis_chat_with_twin` — RAG-grounded Q&A with the Twin's voice (`full_read` only) ## Setup wizard The fastest path: open `https://aigis.bio/view//mcp` (replace `` with the slug from the share URL the candidate sent you). The wizard auto-detects your OS, shows the right config-file path, and gives you copy-pasteable snippets for Claude Desktop, Claude Code, Claude.ai, and generic MCP clients. Paste the bearer token where it says `` and you're done. If you'd rather wire it up manually, the per-client steps below are the same content the wizard generates. ## Setup (manual) You'll need: 1. The share URL the candidate sent you (looks like `https://aigis.bio/view/`). 2. The MCP token (looks like `aigis_sk_<32-character-string>`). The MCP server lives at `https://mcp.aigis.bio` (production) or `https://mcp.staging.aigis.bio` (staging). ### Claude Desktop 1. Open `~/Library/Application Support/Claude/claude_desktop_config.json` (macOS) or `%APPDATA%/Claude/claude_desktop_config.json` (Windows). Create the file if it doesn't exist. 2. Add an `mcpServers` block: ```json { "mcpServers": { "aigis-twin": { "url": "https://mcp.aigis.bio", "headers": { "Authorization": "Bearer aigis_sk_" } } } } ``` 3. Restart Claude Desktop. The Twin's tools appear in the tool drawer. ### Claude Code ```bash claude mcp add --transport http aigis-twin https://mcp.aigis.bio \ --header "Authorization: Bearer aigis_sk_" ``` Restart `claude` or run `claude mcp list` to verify the server registered. ### Claude.ai (web) 1. Settings → Connectors → Add custom connector. 2. URL: `https://mcp.aigis.bio` 3. Authentication: Custom header `Authorization: Bearer aigis_sk_` 4. Save. The Twin's tools become available in your chat. ### Custom agent (any MCP-compatible client) Send `Authorization: Bearer aigis_sk_` on every request to `https://mcp.aigis.bio`. Streamable HTTP and SSE transports are both supported. ## Verifying the connection Ask Claude something like "what tools does this candidate's Twin support?" — the agent should call `aigis_get_twin_capabilities` and return the read-only tool list. If you see `aigis_score_against_jd` in the list, you're wired up correctly. For a full assessment, paste a job description and ask "score this candidate against the JD." The agent will call `aigis_extract_jd_requirements` then `aigis_score_against_jd` and surface a categorical match (`strong` / `partial` / `insufficient`) per requirement with cited verbatim evidence. **A note on AI-assisted screening:** you are interacting with an AI system — the candidate's AI Twin — and the tools say so in their descriptions (an EU AI Act transparency requirement). `score_against_jd` output is assistive analysis to support your review; it must not be the sole basis for a hiring decision, and if you screen candidates with it you are responsible for informing them that AI is used in your process. ## Rate limits Calls are rate-limited per token: - **Heavy** (LLM-driven: `score_against_jd`, `prepare_for_role`, `chat_with_twin`, `ask_candidate`, `get_narrative_anchors`) — 12 calls per 60 seconds. - **Light** (everything else) — 60 calls per 60 seconds. When rate-limited, the response is a structured error with `retry_after_seconds` so your agent can back off automatically. See [Rate Limits](/help?tab=agents-mcp&article=rate-limits) for details. ## Token security - Treat the token like a password. Never paste it into a Slack channel, share screenshots that include it, or commit it to git. - The candidate can revoke the token at any time from their sharing settings. Revocation is effective immediately on the next call. - If you accidentally leak the token, ask the candidate to revoke and reissue. ## Troubleshooting **`unauthorized` error on every call:** the token may be wrong, revoked, or expired. Check it ends with the `aigis_sk_` prefix and you copied the whole string. Ask the candidate to verify the link is still active in their dashboard. **`insufficient_scope` on a tool you expected to work:** that tool requires a scope your token doesn't have. Most likely you tried to call a write tool (`submit_journal`, `submit_artifact`) — share-link MCP tokens are read-only by design, write tools require OAuth. **`tool_not_in_preset` error:** the candidate scoped this share-link to a narrower preset than `full_read`. The error body includes the current preset and the allowed-tool list — fall back to a tool that's covered, or ask the candidate for a wider preset. **`rate_limited` error:** you hit the per-token cap. The error body includes `retry_after_seconds`; wait that long and retry. If you're scoring many candidates in a batch, space the calls or reduce concurrency. **Tool returns `no_evidence`:** the candidate's KB doesn't have direct evidence for what you asked. This is intentional — the Twin marks gaps explicitly rather than imputing weak matches. Ask a different question or use `aigis_query_knowledge` to see what's actually indexed. **Connection refused or 404:** double-check the URL is `https://mcp.aigis.bio` (not `api.aigis.bio` — that's the REST API for the candidate dashboard). Production: `mcp.aigis.bio`. Staging: `mcp.staging.aigis.bio`. For a deeper reference on the available tools and their contracts, see the [public Aigis MCP client skill](https://aigis.bio/skills/aigis-mcp-client.md) — drop it into Claude Code's `~/.claude/skills/` directory or upload to Claude.ai's Skills settings to teach the agent how to use the Twin effectively.