Pursers client

pursers-client is the asynchronous Python client for a Pursers Central MCP service. It provides BoardClient for authenticated board, ticket, memory, state, and event operations used by Pursers runtimes and automation tools.

Applications supply the Central URL, bearer credential, board ID, and seat identity. The client does not start Central, create credentials, or manage operator configuration.

Zed stdio relay

pursers-client installs pursers-mcp, a credential-safe stdio MCP server for Zed and other local MCP hosts:

uvx --from pursers-client==<VERSION> pursers-mcp \
  --central-url http://127.0.0.1:<PORT> \
  --board <BOARD_ID> \
  --token-file /PATH/TO/credential.jwt

Use --ca-file /PATH/TO/private-ca.pem when Central uses a private TLS CA. --central-url accepts either the Central origin or its /mcp endpoint. The token file is read without logging its contents, checked again before every upstream connection generation, and re-read after an HTTP 401. Keep it private (mode 0600). The relay writes only MCP JSON-RPC frames to stdout; connection diagnostics go to stderr and are also exposed as a tool error after the MCP handshake, so Zed can show a missing credential or unreachable Central.

Zed 1.20.2 negotiates MCP 2025-11-25. The relay terminates that local protocol and opens a separate MCP 2026-07-28 connection to Central. It forwards the principal's Central tool names, descriptions, schemas, and call results unchanged. By default it exposes the 18 read/create/annotate, question/human-request, evidence, and resumable-watch tools used by the Zed flows; pass --tools all to expose every tool authorized for the principal. Wait calls are capped at 50 seconds, below Zed's 60-second default, and their upstream cursor result is preserved for the next call. Each request owns an independent upstream SDK context, so a long-running wait cannot block another call and credential rotation cannot cancel an in-flight request.

The relay also publishes five prompts that Zed exposes as slash commands: board, create <summary>, watch, evidence <ticket-id>, and answer <ticket-id-and-answer>. Their instructions keep output bounded, put stable IDs first, and avoid fleet-wide dumps.