Skip to content

Agent integrations

Agent integrations let a coding agent search the Hansard archive without requiring direct access to archive files. The recommended setup installs an MCP configuration when the client exposes a compatible configuration surface, plus the best native command or skill surface available to that agent.

  1. Initialize Hansard and import at least one source.

  2. Add Hansard to the target agent.

    Terminal window
    hansard integrations add-to codex

    Replace codex with claude, antigravity, devin, opencode, or grok.

  3. Restart or reload the target agent so it reads the updated configuration.

  4. Ask the agent to recall a focused topic or continue from a prior session.

Rerun the same add-to command after upgrading Hansard when generated recall instructions need to be refreshed.

Provider memory can contain sensitive project conventions, paths, account context, and durable instructions. It is therefore excluded from agent integrations by default. Enable it explicitly for a trusted agent:

Terminal window
hansard integrations add-to codex --memories

The flag installs targeted memory-retrieval instructions and configures the MCP server to expose the read-only search_memories and get_memory tools. It does not let the agent create, edit, merge, or delete memories.

Local Codex and Claude Code sessions can discover and message each other through the same-project mode selected by default during new setup. Repair or change the managed integration with:

Terminal window
hansard peers setup

This enables the recommended same-project mode and adds messaging tools to Hansard’s existing Codex and Claude MCP entries. Existing sessions must restart once. No skill or manual link command is required. See Message live agent sessions for the drawer, CLI, scope, delivery, and safety behavior.

Codex and Claude Code can launch bounded same-project helpers through the other default-on collaboration setting:

Terminal window
hansard agents setup

This also enables same-project messaging, then adds helper tools to Hansard’s existing Codex and Claude MCP entries. Existing sessions must restart once. The init picker can uncheck both collaboration options, and unattended setup can use --no-agent-collaboration. See Delegate to helper agents for supported harnesses, nested ownership, queueing, safety limits, and the end-to-end transcript.

Native surfaces vary because agent clients expose different extension points.

Target Installed surface
Codex MCP configuration plus recall, continue-from, and Deep Recall skills.
Claude MCP configuration plus slash commands and matching skills.
Antigravity MCP configuration plus recall and continue-from skills.
Devin MCP configuration plus recall and continue-from skills.
OpenCode MCP configuration plus recall and continue-from commands.
Grok Build MCP configuration plus recall and continue-from surfaces compatible with Grok Build.

Each target also receives its native memory-retrieval surface when --memories is present.

Use integrations recall when the MCP configuration is managed separately:

Terminal window
hansard integrations recall codex

The current live targets are codex, claude, antigravity, devin, opencode, and grok.

Clients with a manually managed MCP configuration can launch the recall server over standard input and output:

Terminal window
hansard mcp serve

The server exposes search_past_sessions by default. To opt a manually managed client into memory retrieval, launch hansard mcp serve --memories instead. The memory-enabled server also exposes the read-only search_memories and get_memory tools. A manually managed Codex or Claude client can use hansard mcp serve --delegation; delegation implies the peer messaging tools. Generated integrations call the installed hansard executable, so a global installation must be available on PATH unless the client configuration supplies another executable path.

Finds conversations relevant to a focused topic, then reads a bounded latest-message window before reporting the result.

Generated instructions tell the agent to infer a focused query rather than passing an entire user request directly to archive search. Archive-dependent claims should remain tied to the selected session IDs.

Memory-enabled instructions tell the agent to search proactively when work depends on a prior preference, provider capability, durable project convention, or earlier decision. They do not inject the full unified memory archive into every prompt.