v1.5.2

Fixed

• SSE servers that retired the legacy transport. Cloudflare's MCP servers (and others moving
  off HTTP+SSE) now answer the SSE stream with 410 Gone, which left the server stuck on
  "Unreachable · Non-200 status code (410)" and gave no hint about the fix. When an SSE server
  refuses the stream (404/405/410), the switchboard now retries Streamable HTTP at the same URL
  and switches the server to HTTP once that connects. OAuth servers go to Needs auth as usual.
  If HTTP fails too, the error says to change the transport and URL (usually /sse → /mcp).

v1.5.1

Fixed

• Claude connectors behind chained proxies. Behind Cloudflare → Caddy (or any proxy hop that
  speaks plain HTTP), the agent OAuth metadata advertised http:// URLs, because the inner proxy
  reports its own scheme, and Claude refused to sign in. A PUBLIC_URL that is https:// now
  settles the scheme for requests to its own host, so the common one-hostname setup works without
  also setting MCP_PUBLIC_URL.

Changed

• Clearer env var docs. The README now splits the variables into the admin UI's and the agent
  endpoint's, spells out which OAuth direction PUBLIC_URL and MCP_PUBLIC_URL each serve, lists
  typical setups, and adds a Behind a reverse proxy section with a curl check for the metadata
  Claude reads. The Claude connector steps now match the current dialog (full /mcp/<slug> URL,
  dynamic registration, CIMD unsupported).

v1.5.0

Added

• OAuth for agents — connect Claude Desktop and claude.ai. Custom connectors in Claude Desktop
  and claude.ai can't send a static Authorization header, so they couldn't reach
  /mcp/<agent-slug> at all. The switchboard is now its own OAuth 2.1 authorization server for its
  agent endpoints (protected-resource and authorization-server metadata, dynamic client
  registration, PKCE, refresh-token rotation): paste the agent's endpoint URL into Add custom
  connector and approve the connection with a one-time pairing code generated on the Agents
  page. Pairing codes keep the agent's long-lived token off a web page anyone can link to. Approved
  clients are listed per agent under OAuth connections and can be revoked; static bearer tokens
  keep working unchanged. The endpoint must be reachable over public HTTPS, since Claude's
  connectors connect from Anthropic's cloud.
• CORS on /mcp/<agent-slug>, so browser-based clients such as MCP Inspector can read the
  session id and the OAuth challenge.

v1.4.3

Fixed

• OAuth no longer breaks for good when PUBLIC_URL changes. A dynamically registered OAuth
  client is bound to the redirect URI it was registered with, so after PUBLIC_URL changed the
  stored registration kept sending the old URI and providers rejected every attempt with
  "Invalid redirect URI … does not match any registered URI for this client" (Linear's wording) —
  with no way out from the UI. A registration whose redirect URI no longer matches is now discarded
  and re-created on the next authorization, and a redirect-URI error explains that the provider
  wants an https or loopback PUBLIC_URL.

v1.4.2

Fixed

• Idle agent sessions are less likely to be dropped by a reverse proxy. The MCP SDK was updated
  to 1.30.0, whose Streamable HTTP server transport now emits SSE keep-alive comment frames (and
  fixes the keep-alive timer's lifecycle) — an agent that sits idle between tool calls no longer
  looks like a dead connection to an intermediary that times out silent streams, which showed up as
  a session vanishing mid-conversation. The same bump tightens Content-Type validation to parse
  the media type instead of substring-matching it, and bounds the read buffer for stdio servers that
  emit very large responses.

v1.4.1

Fixed

• git in the Docker image. stdio MCP servers installed from a git+https://… spec (e.g. via
  pipx run / uvx) failed with Cannot find command 'git' because the runtime image only had
  python and pipx. git and ca-certificates are now installed alongside them.

v1.4.0

Added

• Lean mode per agent. An agent with 5–10 enabled servers can receive 100+ tool definitions —
  tens of thousands of context tokens paid on every session for schemas it mostly never uses. A new
  per-agent Tool exposure toggle (Agents page, default Full) switches an agent to Lean: its
  tools/list shrinks to a constant-size set of meta-tools — switchboard__search_tools (ranked
  lexical search over its enabled catalog), switchboard__describe_tools (full descriptions with
  input/output shapes rendered as compact TypeScript instead of JSON Schema), and
  switchboard__call_tool (invoke any enabled tool by namespaced name) — so tools are discovered on
  demand via search → describe → call. The switch matrix still governs what is searchable and
  callable, typo'd names come back with suggestions, the Logs page attributes wrapped calls to the
  real upstream tool, and toggling the mode updates live sessions immediately.

• Agents can register MCP servers themselves. Discovering a useful server used to mean relaying
  its config to a human to paste into the UI. An agent granted the new manager role (a per-agent
  toggle on the Agents page, off by default) gets switchboard__add_server — paste the same
  claude mcp add … line or mcpServers JSON found in a README and the server is created, enabled,
  and its tools appear in the same live session — plus switchboard__assign_server,
  switchboard__list_agents and switchboard__server_status to manage the matrix and self-debug.
  Local stdio servers are the deliberate exception: they execute a command on your machine, so no
  agent can create one directly, whatever its role.

• A request/approval queue keeps humans in the loop. Every agent (manager or not) gets
  switchboard__request_server: it files the ask — with a pasted config or just a reason — and
  returns immediately. Pending requests show as a badge in the UI with a review panel that renders
  stdio commands in a "this will execute on your machine" warning block; approving one creates the
  server through the same path as the UI and pushes the new tools into the requesting agent's
  still-open session. Denials carry an optional note the agent reads back with
  switchboard__request_status. Stored request configs are encrypted at rest like every other
  secret, and a Settings toggle hides the request tools entirely if you'd rather agents couldn't
  ask. Servers created by agents show an "added by" chip for provenance, and because meta-tool
  calls are ordinary MCP requests, every management action lands in the request logs.

Changed

• Calls to a server an agent lacks now fail exactly like calls to a server that doesn't exist.
  The two cases used to return different errors ("not enabled for this agent" vs "unknown server"),
  which let any agent probe which server slugs were configured globally. Both now return the same
  Unknown server error, so the switch matrix no longer leaks the existence of servers it hides.

Fixed

• Request logs no longer attribute switchboard__* meta-tool calls to a phantom "switchboard"
  server. Management and request meta-tools were logged with switchboard as their upstream
  server — a server that doesn't exist — polluting the Logs page's server filter. They now log with
  no upstream server, like the other switchboard-native rows.

v1.3.0

Added

• Oh My Pi connection snippet. The connect dialog only covered Claude Code, Codex and raw JSON,
  so Oh My Pi users had to hand-translate the URL and token into its /mcp add syntax. A fourth tab
  now gives a ready-to-paste slash command.

v1.2.0

Added

• MCP_PORT puts the agent endpoint on its own port. The UI, the REST API and the agent
  endpoint all shared one port, so exposing /mcp/<agent> beyond the LAN meant exposing the admin
  UI with it. Set MCP_PORT and the same process opens a second listener that serves only
  /mcp/<agent-slug> — you can then firewall, tunnel or reverse-proxy that port on its own while
  the UI stays on the trusted interface. Both listeners share the same database, upstream
  connections and sessions. HOST / MCP_HOST bind each listener to a specific interface, and
  MCP_PUBLIC_URL overrides the base URL shown in the connection snippets when the endpoint is
  reached through a proxy or tunnel.
• Request logs. A new Logs page records every JSON-RPC request an agent sent through the
  switchboard and the response sent back, so "the agent says the tool failed" stops being a
  guessing game. Each request is one line — time, agent, upstream server, tool name, arguments
  preview, duration — that expands to the exact frames on the wire. Filter by agent, by server,
  by method, or by errors / slow (>1s) / in-flight, and search across tool names, summaries and
  error messages. Errors, slow calls and requests still in flight are distinguishable by shape,
  not only colour. Headline counts for the last 24h (volume, error rate, median, p95) sit above
  the list.
• The log tails live. New requests appear as they arrive over a server-sent-events stream,
  including long-running calls, which show up as in-flight the moment they start and update in
  place when they finish. Some reverse proxies buffer or drop event streams; when the stream
  can't be established the page falls back to polling and says so.
• Log retention and payload capture are configurable in Settings → Logs. Requests are kept
  for 48 hours by default and pruned automatically; shortening the window prunes immediately.
  Full request/response payloads are recorded by default (capped at 32 KB per frame, with the
  true size still reported) — turning capture off keeps the metadata but stores no frames, which
  matters because tool arguments pass through unredacted.

v1.1.0

Added

• Light / dark / system theme switcher. The dark palette was previously locked to the OS
  prefers-color-scheme setting with no way to override it. The choice now persists per
  browser and, in system mode, still follows the OS live. The segmented picker sits in the
  sidebar footer and in Settings → General (the sidebar is hidden below the sm breakpoint).
• Published to npm as @cmer/mcp-switchboard, runnable with npx @cmer/mcp-switchboard.
  The package bundles the built UI alongside the compiled server, so trying the switchboard no
  longer requires cloning the repo or installing Docker. Docker remains the better fit for a
  long-running install.

Changed

• The data directory now defaults to ~/.config/mcp-switchboard (respecting
  XDG_CONFIG_HOME) instead of ./data relative to the working directory. The old default
  meant the switchboard you got depended on which directory you launched it from — harmless
  under Docker, which pins DATA_DIR, but a real hazard for npx, where it would scatter
  databases and encryption keys across the filesystem. Existing installs keep working: if a
  cwd-relative data/switchboard.db is found and the new location is empty, it is used and a
  warning points at the new path. DATA_DIR still overrides everything.
• The data directory is now created 0700. It holds the encryption key and every stored OAuth
  token, and was previously world-readable (the key file itself was already 0600).