context7

Patch Changes

• 2affada: ctx7 setup now properly supports --antigravity, installing skills to .agent/skills, a GEMINI.md rule section (Antigravity reads Gemini-family config), and MCP config to Antigravity 2.0's documented global path ~/.gemini/config/mcp_config.json (with httpUrl for HTTP, matching the Gemini convention). Antigravity has no documented project-level MCP file, so setup --antigravity --project --mcp writes to the global location. Also removes the --universal flag from setup, which was advertised but silently ignored — it never propagated through agent selection, so passing it (e.g. setup --cli --universal --project) caused setup to fall back to auto-detection and write to the wrong directory.
• 268f52f: ctx7 setup --api-key <KEY> (without --cli, --mcp, or -y) now prompts to choose between MCP server and CLI + Skills modes. Previously, passing --api-key short-circuited to MCP, locking users out of the CLI + Skills option even though that mode also accepts an API key. Explicit --mcp / --cli / --stdio / --oauth / -y still skip the prompt as before.
• 2e97dae: Add deprecation warning to skill commands

Minor Changes

• 1fb2d42: Add multi-tenant Microsoft Entra ID validation for MCP tokens. The server now detects inbound Entra v2 tokens by issuer pattern, fetches per-teamspace configuration (tenantId, audience, requiredScope) from the Context7 app, and verifies the token against the matching tenant's JWKS, enforcing the required scope claim when configured. User resolution happens downstream in the Context7 app against a pre-provisioned user mapping table — the MCP server only validates. Per-tenant JWKS cache and a 5-minute in-memory config cache keyed by JWT audience reduce overhead under load.

Minor Changes

• f91b40c: Initial release. Adds an official Context7 extension for the pi coding agent — registers resolve-library-id and query-docs tools, ships the context7-docs skill, and exposes a /c7-docs slash command. Wire format, error messages, and tool descriptions are copied verbatim from @upstash/context7-mcp so pi and MCP clients give the LLM identical instructions and output. Self-contained — no Context7 runtime dependencies. Works out of the box at IP-based rate limits; set CONTEXT7_API_KEY for the higher tier. Install with pi install npm:@upstash/context7-pi.

Patch Changes

• 7cacc94: Add --json flag to ctx7 skills list for machine-parseable output. Emits { skills: [{ name, path, source }] } where path is absolute and source is the agent type (universal, claude, cursor, antigravity). Matches the existing --json pattern on ctx7 library and ctx7 docs.

Major Changes

• af6a7b5: Convert the stateless MCP implementation to a stateful one using Redis for session management.

Patch Changes

• 3d73145: Reduce Redis writes on refresh by checking the remaining TTL first and only issuing EXPIRE when the session is within one day of expiry.

Minor Changes

• 34fda7d: Prompt anonymous users to sign in. After the backend signals (via the X-Context7-Auth-Prompt: 1 response header on /v2/libs/search or /v2/context) that an anonymous client has crossed the per-IP threshold, the MCP server appends a one-time sign-in invitation to the tool result.
  • Both stdio and HTTP transports surface the same nudge: a tool-result notice asking the assistant to run npx ctx7 setup --<client> --mcp -y (with --stdio appended when the MCP server is running on stdio) after explicit user confirmation. The CLI handles OAuth and writes credentials into the MCP client's config; the user restarts their MCP server / editor to pick up the new credentials.
  • Detects the calling client from X-Context7-Client-IDE / User-Agent and selects the matching CLI flag (--cursor, --claude, --codex, --opencode, --gemini); falls back to interactive setup when unknown.
  • HTTP transport remains stateless — the threshold is tracked by the backend (per-IP, 24h TTL), the MCP server only reacts to the signal.

Patch Changes

• 6c71e4d: Handle malformed MCP config files gracefully during ctx7 remove agent detection. Previously, an unparseable JSON config at any agent's well-known path (e.g. a hand-edited ~/.claude.json) would crash the command with an unhandled SyntaxError before it could do anything. The detector now skips the offending file and logs a warning naming the path and parse error so the user can fix it, while detection continues for the remaining agents.
• 4056850: Respect CLAUDE_CONFIG_DIR env var when resolving Claude Code's global config, rules, skills, and detection paths

Patch Changes

• 187287c: Accept hallucinated argument names on tools/call requests by rewriting them to the canonical names before validation. userQuery and question are mapped to query on either tool; on query-docs, context7CompatibleLibraryID, libraryID, and libraryName are mapped to libraryId. Some LLM clients produce these alternative names — likely echoing phrasing from each tool's description — and previously triggered Invalid input: expected string, received undefined errors. libraryName is only rewritten on query-docs calls because it is the canonical arg for resolve-library-id. Tool input schemas published via tools/list are unchanged: canonical names remain the documented required fields, the rewrite is purely a server-side compatibility shim that runs only on tools/call and only when the canonical key is absent.
• 78b9826: Exit the stdio MCP server when the parent process closes its stdio. Previously, if the parent (e.g. Claude Code) was force-killed shortly after a tool call, an idle undici keep-alive socket to the Context7 API would keep libuv's event loop alive past stdin EOF, leaving an orphaned node process that consumed memory until the kernel tore the socket down (which on Cloudflare-fronted endpoints can take hours). The server now listens for end/close on stdin and SIGHUP and exits cleanly. Fixes #2542.

Patch Changes

• d0e4a48: Create a fresh McpServer per HTTP request. Sharing one across requests let any concurrent transport.close clear the shared Protocol._transport, which broke sendNotification for in-flight long-running tool calls.
• 1aa3430: Remove research mode entirely from the MCP server and CLI. The query-docs MCP tool no longer accepts or forwards a researchMode parameter, and the CLI no longer exposes a --research flag on ctx7 docs.

Patch Changes

• 772da3a: Stream MCP tool responses over SSE so HTTP headers flush before client fetch timeouts. Switching enableJsonResponse to false makes the SDK return the HTTP response synchronously after request validation, so headers are sent in milliseconds instead of being buffered until the tool completes. This fixes clients that cap the underlying fetch waiting for headers (e.g., Claude Code's 60s wrapFetchWithTimeout).

Patch Changes

• 8274bd0: Add missing tool annotations
• ff6c1be: Remove the researchMode parameter from the query-docs tool's input schema. The underlying API still supports research mode, but several MCP clients hit per-request timeouts (60s defaults) on long-running research calls in ways that can't always be solved server-side. Hiding the parameter prevents agents from invoking it through MCP until the timeout story is reliable across clients.

Patch Changes

• 1b0c211: Add endpoint for OpenAI Apps SDK domain verification.

Minor Changes

• 17b864f: Expose research mode through the MCP researchMode tool and the CLI docs --research flag for deep, agent-driven documentation answers.

Patch Changes

• 4feee15: Add CLI update notifications and a new ctx7 upgrade command. The CLI now checks for newer versions with cached state, shows a non-blocking notice before interactive commands, and provides safer upgrade guidance across npm, pnpm, bun, and ephemeral runner setups.
• f056b14: Add ctx7 remove as the cleanup counterpart to ctx7 setup, with safer detection and removal behavior. The command now prompts only for agents with actual Context7 artifacts, preserves non-Context7 MCP configuration when removing entries, and includes stronger test coverage for JSON and TOML cleanup.

Patch Changes

• 3f6e310: Fix skill installation path validation on Windows so valid files inside the target directory are not rejected due to backslash-separated resolved paths.

Patch Changes

• 00833f9: Preserve Node's default trusted CAs when NODE_EXTRA_CA_CERTS is configured, and add a regression test for custom CA loading.

Patch Changes

• 33f2338: Add Codex-specific CLI setup guidance so generated rules and the installed find-docs skill tell Codex to rerun Context7 CLI requests outside the default sandbox after DNS or network failures.

Patch Changes

• bc8eaf1: Add --all-agents and --yes support to ctx7 skills install for non-interactive multi-agent installs.

Patch Changes

• fb29170: Add Gemini CLI support to setup command
• 89d4862: Use GITHUB_TOKEN/GH_TOKEN or gh CLI auth for skill downloads to avoid GitHub API rate limits and support private repos
• 8322879: Improve resolve libryar id tool prompt to provide the libraryName query with proper format

Patch Changes

• 6961bdd: Allow re-selecting already configured agents in ctx7 setup and overwrite existing MCP config entries instead of skipping them. Fix TOML replacement to correctly handle sub-sections and prevent whitespace drift on repeated runs.

Patch Changes

• a667712: Update search filter warning

• d739f9b: Fix OpenCode MCP setup to resolve all config file variants (opencode.json, opencode.jsonc, .opencode.json, .opencode.jsonc)

• 4f13168: Install rules alongside skills in ctx7 setup for better trigger rates

  • CLI setup now installs a rule file for each agent (previously only installed the skill)
  • Rule content fetched from GitHub, with agent-specific formatting (alwaysApply for Cursor)
  • Updated find-docs skill description for higher invocation rates (66% -> 98%)
  • Added Codex agent support with AGENTS.md append
  • OpenCode now writes to AGENTS.md instead of .opencode/rules/
  • Selective rule content with explicit when-to-use/when-not-to-use guidance

• c3c2647: Use ~/.agents/skills instead of ~/.config/agents/skills for global universal skill installs