Provider and feature support
This reference describes implemented Hansard support, reviewed on September 29, 2026. Availability on a computer also depends on installed clients, enabled integrations, credentials, and access to the original conversation.
A harness runs the agent, such as Codex or OpenCode. A model provider supplies inference. Selecting a model, switching accounts, importing history, and attaching Hansard tools are separate capabilities.
The import coverage inventory lists every registered source and format by product, including browser routes, verification gaps, and unsupported products.
Yes means Hansard implements the feature. Opt-in means the feature requires an additional setting. No managed integration does not mean the harness itself lacks MCP support. A compatible client can be configured manually with the shared Hansard tool server. Coordination additionally needs an identified conversation.
Managed agent tools
Section titled “Managed agent tools”Managed setup now covers 16 clients. Other local clients can use the same MCP server or the CLI binding path below; the installer list is not a provider allowlist for status or coordination. Support follows the harness and session identity, independently of the chosen model. Instructions ask the model to use tools; they cannot guarantee model compliance.
Bound means the tools are installed and enabled by default, but the client
must supply its actual conversation ID through hansard session bind first.
The returned token is passed on every status, peer, board, or helper call.
Recall and memory search need no conversation binding.
Harness (add-to target) |
Conversation search/fetch | All-provider memory search/fetch | Status reports | Peer messages and boards | Launch helpers as a parent | Receive work as a helper | Automatic identity |
|---|---|---|---|---|---|---|---|
Codex CLI and app (codex) |
Yes | Default on | Default on | Default on | Default on | Yes | Native hooks and per-request metadata |
Claude Code (claude) |
Yes | Default on | Default on | Default on | Default on | Yes | Native hooks |
Grok Build (grok) |
Yes | Default on | Default on | Default on | Default on | Yes | Native hooks |
Antigravity CLI (antigravity) |
Yes | Default on | Default on | Default on | Default on | Yes | Pre-invocation hook |
Antigravity app and IDE (antigravity) |
Yes | Default on | Default on | Default on | Default on | No | Pre-invocation hook |
Devin CLI (devin) |
Yes | Default on | Default on | Default on | Default on | No | Native hooks |
OpenCode (opencode) |
Yes | Default on | Default on | Default on | Default on | Yes | Native plugin |
Factory Droid (factory) |
Yes | Default on | Default on | Default on | Default on | No | Native hooks |
Augment CLI (augment) |
Yes | Default on | Default on | Default on | Default on | Yes | SessionStart hook |
Cursor app and CLI (cursor) |
Yes | Default on | Default on | Default on | Default on | No | sessionStart hook |
Cline CLI (cline) |
Yes | Default on | Bound | Bound | Bound | Yes | Explicit binding |
GitHub Copilot CLI (copilot) |
Yes | Default on | Bound | Bound | Bound | Yes | Explicit binding |
Amp (amp) |
Yes | Default on | Bound | Bound | Bound | Yes | Explicit binding |
Kimi Code (kimi) |
Yes | Default on | Bound | Bound | Bound | Yes | Explicit binding |
Kiro (kiro) |
Yes | Default on | Bound | Bound | Bound | Yes | Explicit binding |
Zed (zed) |
Yes | Default on | Bound | Bound | Bound | No | Explicit binding |
Claude Desktop (claude-desktop) |
Yes | Default on | Bound | Bound | Bound | No | Explicit binding; requires an accessible native ID |
Status tools are installed even when messaging is off. Messaging, boards, and
helper launches default to the same project. Messaging and boards follow
orchestrate.messaging.mode; parent helper tools follow
orchestrate.delegation.mode. Turning messaging off also disables delegation.
Memory retrieval is enabled by default and independent of both settings.
--no-memories disables memory tools for an integration.
Run hansard integrations sync to update existing entries, then reload clients.
Sync preserves explicit off settings.
Codex and Claude Code also receive deep-recall and unify-memories workflows.
Codex, Claude Code, Cursor, Grok Build, Antigravity, Devin, OpenCode, and Factory
receive recall, continue-from, and recall-memories by default. OpenCode uses commands;
Claude Code receives commands and skills. The other managed clients use MCP
tools directly without generated native skills.
Native sessions and Hansard launches receive session guidance where hooks or
plugins are available. Existing clients need a reload. Codex may require
approving hooks through /hooks. Grok’s tool-call hook supplies identity.
The Cursor app and the Cursor CLI share one integration target because both
read ~/.cursor. Augment’s managed
configuration targets its CLI; Cline’s IDE configuration is separate from the
managed CLI file. Claude Desktop’s local MCP entry does not establish a Cowork
sandbox integration. A local SDK run must load MCP configuration and identity;
importing an SDK transcript alone does not attach tools.
Peer messages always have a durable inbox. Live previews and work delivery
depend on host transport. Added MCP installers do not create native push or
resume adapters. Clients without those transports check peer_inbox; a message
does not start another agent process. Boards support project discussions and
explicitly joined shared topics across projects.
Other clients and remaining gaps
Section titled “Other clients and remaining gaps”Export the default tool invocation for any compatible local stdio MCP client:
hansard integrations configThe output uses the standard mcpServers JSON shape. Translate the entry into
the client’s native format when it differs. For example, Vibe uses TOML MCP
server arrays, and current Crush uses its executable crushrc configuration.
Hansard does not yet manage those files or native lifecycle hooks.
For a client without native identity integration, obtain its actual current conversation ID and bind it locally:
hansard session bind --provider moonshot --native-session-id <native-id> --cwd <project>moonshot is the archive provider ID for Kimi Code. Use the corresponding archive
provider ID for another client, such as cursor, amp, mistral, or openhands.
The binding path accepts every provider in Hansard’s source catalog. Keep the
returned token in that conversation and pass it as session_token on MCP calls,
or --session-token on hansard session report. Never invent an ID, copy another
conversation’s token, or put one conversation’s token in shared global MCP config.
If the client does not expose its native ID, automatic coordination remains
unavailable; read-only recall and memory tools still work.
| Remaining surface | Available path | Missing managed capability |
|---|---|---|
| Other local clients with stdio MCP | Export the common invocation, configure the client, then bind its native ID | Native config writer, lifecycle identity hooks, and client runtime verification |
| Local clients with shell access but no MCP | CLI archive/memory search and explicit session binding; report through the CLI | Automatic tool attachment and lifecycle instructions |
| Other IDE extensions, including Cline IDE | Configure the extension’s own MCP settings, where supported | Extension-specific settings discovery and automatic identity |
| Browser chats and hosted agents | Archive imports and memory retrieval from a connected local agent | A reachable authenticated MCP transport or supported host connector; local stdio cannot run inside a remote browser service |
| Cowork and other remote sandboxes | Existing archive intake | Sandbox-accessible MCP configuration and per-conversation identity |
| Gemini CLI | Historical archive import | Intentionally excluded from new live setup; use Antigravity |
| Providers without a launch adapter or request-router client | Their own native application controls | Separate launch/resume/router adapters; installing MCP does not implement them |
The added config writers are checked against documented provider formats and synthetic integration tests, including preservation of foreign settings, refresh, opt-outs, and hook output shapes. These checks do not establish that every client version has been exercised in a live conversation.
Complete MCP tool inventory
Section titled “Complete MCP tool inventory”The main server exposes up to 33 tools. Tool groups are shared across models; the table above describes which integrations install and identify them.
| Group | Tools | Availability |
|---|---|---|
| Conversation recall, 2 | search_past_sessions, get_past_session |
Every MCP server invocation |
| Memories, 2 | search_memories, get_memory |
Default; --no-memories omits them |
| Session status, 2 | report_status, session_context |
--status, --messaging, or --delegation |
| Peer messaging, 7 | list_peers, send_message, ask_session, reply_message, peer_inbox, peer_read, set_peer_messaging_muted |
Messaging enabled |
| Boards, 14 | board_list, board_create, board_join, board_leave, board_members, board_channels, board_create_channel, board_post, board_threads, board_thread, board_read, board_search, board_subscribe, board_updates |
Messaging enabled |
| Helpers, 6 | list_agent_targets, spawn_agent, list_agents, wait_agents, message_agent, cancel_agent |
Delegation enabled; default for all managed integrations |
Memory tools read archived content. They do not edit files or cloud memories.
unify-memories is a workflow, not an additional MCP tool. Project coordinators
receive an additional Hansard MCP connection with memory, messaging, and helper
tools enabled. Ordinary agent integrations also include memory retrieval
unless explicitly disabled.
Launches, continuation, and model controls
Section titled “Launches, continuation, and model controls”These operations run on the current computer. They need an installed harness and an existing working directory. Native resume and fork also need the original conversation to be available to that harness. Conversation controls hide unavailable harnesses on the selected device. Settings → Handoff & helpers lists installed harnesses with enable switches and folds the rest of the launch inventory under a disclosure. Detection requires the executable; historical data and terminal wrappers alone do not qualify. Installation does not imply authentication or a compatible CLI version.
| Harness | Start a new conversation or receive a handoff | Resume original conversation | Native fork | Select model | Select effort | Live interaction in Hansard | Project coordinator |
|---|---|---|---|---|---|---|---|
| Codex CLI, app, SDK history | Yes | Yes | Yes | Yes | Yes | Yes, persistent transport | Yes |
| Claude Code CLI and SDK history | Yes | Yes | Yes | Yes | Yes | Yes, persistent transport | Yes |
| Grok Build | Yes | Yes | No | Yes | No separate control | ACP approvals, questions, cancellation; no steering | No |
| Antigravity CLI | Yes | Yes | No | Yes | Yes | No persistent interaction transport | No |
| Antigravity app and IDE | No native launch adapter | No | No | No Hansard launch control | No | No | No |
| Devin CLI | Yes | Yes | No | ACP model override | No separate control | ACP approvals, questions, cancellation; no steering | No |
| Cursor CLI | Yes | ACP-started sessions over ACP; terminal chats through headless cursor-agent -p --resume |
No | Cursor’s model catalog | Per-model effort from Cursor’s catalog | ACP approvals, questions, cancellation; no steering; headless resumes ask for no approvals | No |
| OpenCode | Yes | Yes | No | Yes | No separate control | ACP approvals, questions, cancellation; no steering | No |
| Amp | Yes | Local threads | No | Amp configuration | No | Streaming output and stop; no interactive approvals | No |
| Augment CLI | Yes | Local CLI sessions, ACP loading required | No | ACP model override | No | ACP approvals, questions, cancellation; no steering | No |
| GitHub Copilot CLI | Yes | Local CLI sessions, ACP loading required | No | ACP model override | No | ACP approvals, questions, cancellation; no steering | No |
| Kimi Code | Yes | Local CLI sessions, ACP loading required | No | ACP model override | No | ACP approvals, questions, cancellation; no steering | No |
| Kiro CLI | Yes | CLI and unified local sessions, ACP loading required | No | ACP model override | No | ACP approvals, questions, cancellation; no steering | No |
| Cline CLI | Yes | Local SDK-format sessions, ACP loading required | No | ACP model override | No | ACP approvals, questions, cancellation; no steering | No |
| Mistral Vibe | Yes | Local CLI sessions, ACP loading required | No | ACP model override | No | ACP approvals, questions, cancellation; no steering | No |
| Other imported sources | No native launch adapter | No | No | No Hansard launch control | No | No | No |
The added ACP launchers follow the provider entry points:
Augment (auggie --acp),
Copilot (copilot --acp --stdio),
Kimi (kimi acp),
Kiro (kiro-cli acp),
Cline (cline --acp), and
Vibe (vibe-acp).
Devin always runs over ACP (devin acp), including under the exec
transport preference. The Cursor CLI starts conversations over
ACP (cursor-agent acp, found as
cursor-agent or as the agent link its installer creates). Cursor’s ACP
server opens only the sessions it created, so a conversation started in the
Cursor CLI’s terminal resumes through a headless cursor-agent -p --resume
turn, and a chat missing from its store fails instead of starting over. Model
and effort choices come from Cursor’s parameterized model catalog, and Cursor
saves a model chosen over ACP as its CLI model. Permissions are Agent, Plan,
Ask, or Run Everything (--force); the default keeps the Cursor CLI’s own
approval mode. Resume negotiates session/resume
or session/load; an agent without either capability fails before sending the
prompt. A rejected model override also fails before prompting. No approval-bypass
flags are added. Amp uses --execute --stream-json and threads continue.
These seven launch adapters attach the Hansard MCP server and session identity per run. Memory retrieval and collaboration follow the saved opt-outs; a managed integration explicitly forgotten by the user is not injected. Directly opened sessions still need the managed or manual setup described above. Vibe’s native configuration remains manual; Hansard-launched Vibe sessions receive the tools.
The new transports have synthetic protocol and identity coverage. Native provider authentication and end-to-end behavior require validation with each installed CLI. Factory, IDE-only stores, and browser chats remain outside native launch support. Existing explicit handoff/helper provider lists are preserved; new installations enable all twelve launch targets by default.
A conversation from another source can still supply context to a supported
handoff target. Web chats can be continued through continue-from inside an
existing agent. A synced conversation can start a local continuation; it does
not become a locally resumable native session merely because it was synced.
Model selection uses the harness’s model names and available choices. Hansard does not convert an unsupported native model into a supported one. Fable is a Claude Code model alias, so it follows Claude Code’s capabilities.
Account and request routing
Section titled “Account and request routing”Session pools choose a native account before a session starts. Request routing sends model requests through the managed local router. Router sign-in adds an upstream account to that router; it does not configure every harness to use it.
| Service or harness | Native account profiles and sign-in | Native session pools | Managed request-routing client | Router upstream sign-in | Usage information |
|---|---|---|---|---|---|
| Codex / OpenAI | Yes | Yes | Yes | Yes | Native and router quota capture |
| Claude Code / Anthropic | Yes | Yes | Yes | Yes | Native and router quota capture |
| Grok Build / xAI | Yes | Yes | No | Yes | Native and router quota capture |
| Antigravity / Google | Device account only | No | No | Yes | Native and router quota capture |
| Devin | Yes | Yes | No | Yes | Native and router quota capture |
| Kimi | No managed native profile | No | No | Yes | Router quota capture |
| Kimi.ai | No managed native profile | No | No | Yes, separate login | Router quota capture |
| Meta / Muse | No managed native profile | No | No | Yes | No router quota reader |
| OpenCode | No Hansard account manager | No | No | No dedicated upstream login | Local activity and cost statistics, not verified plan quota |
| Cursor CLI / Cursor | Yes | Yes | No | No | Native plan, included usage, and on-demand spend from the Cursor CLI’s and the Cursor app’s sign-ins; official member or organization usage APIs |
| Factory | No | No | No | No | Official member or organization usage APIs |
| GitHub Copilot | No | No | No | No | Personal or organization usage APIs |
| Browser chats and other harnesses | No managed native profile | No | No | No dedicated upstream login | Imported usage when the source records it |
Codex and Claude Code have automatic account switching, account-order and
usage-based rules, and an opt-in automatic Continue after a usage limit.
Automatic continuation applies to conversations running in Hansard and keeps
the same provider and model. Request routing does not automatically change a
Codex conversation into Claude, Grok, or another provider when quota runs out.
The advanced router supports several upstream providers, but its account and model availability must be checked separately. Hansard’s managed client setup is limited to Codex and Claude Code. Provider-native custom endpoints or model routing in another harness are outside this managed support claim.
Saved browser profiles are available for ChatGPT, Claude.ai, Cursor, Devin, and Grok. They keep browser sign-ins separate; they do not join native account pools or route inference. Connecting an external compatible router lets Hansard inspect and change its routing strategy; that router manages its own accounts.
Archive and memory capture by source
Section titled “Archive and memory capture by source”This table covers every named source family in the source catalog, including cloud variants. Managed tools refers to the first table. Import support does not imply support for launching an agent or attaching tools to it.
Project files only means there is no dedicated provider-memory reader.
Hansard can still capture recognized instruction files from known projects,
such as AGENTS.md and CLAUDE.md, and recover supported archived file edits.
Those shared files are stored as project memories, not as a complete export of
that provider’s private memory system.
Skills are not memories, so memory backup does not capture skill folders. The Skills & plugins view lists skills for the harnesses it supports.
| Source family | Conversation archive input | Memory or instruction capture | Managed tools |
|---|---|---|---|
| Aider | Local Markdown chat history | Project files only | No |
| Amp | Historical local history, JSON exports, explicit cloud import | Guidance files | Yes; explicit binding |
| Antigravity CLI | Local transcript logs, summaries, brain artifacts, usage databases | CLI knowledge and implicit stores; binary content is preserved without text decoding | Yes |
| Antigravity app | Local transcript logs, summaries, brain artifacts, usage databases | App knowledge and implicit stores; binary limitation applies | Yes |
| Antigravity IDE | Local transcript logs, summaries, brain artifacts, usage databases | IDE knowledge and implicit stores; binary limitation applies | Yes |
| Augment | CLI sessions, IDE history, local Cosmos caches, explicit cloud import | Guidelines, rules, commands, migrated context, local Cosmos memory Markdown | Yes for CLI; native SessionStart identity |
| ChatGPT | Account export or browser capture | Manual import of copied saved memories; no automatic private-memory download | No |
| Claude Code | Local CLI and SDK history; explicit cloud and Remote Control import | Global instructions and project memory files | Yes for local Claude Code; cloud imports do not install tools |
| Claude Desktop and Cowork | Desktop metadata and local workspace history | Readable Cowork memory files; underlying Claude Code memories use the Code reader | Desktop MCP only; explicit binding; Cowork sandbox not covered |
| Claude.ai | Account export or browser capture | Memories supplied in the account export; browser chat capture alone does not export memories | No |
| Cline | IDE tasks and CLI / SDK histories | Project files only | Yes for CLI; explicit binding; IDE configured separately |
| CodeBuddy | Local session history | Project files only | No |
| Codex | CLI, app, and SDK histories; explicit cloud tasks | Memory directory, rollout summaries, global instructions | Yes locally; cloud task imports do not install tools |
| Continue | Local histories | Global and project rules and prompts | No |
| Crush | Local histories | Global and project CRUSH.md / crush.md |
No |
| Cursor | Local app stores and Cursor CLI chat and ACP stores; explicit cloud import | Local pending memories and user rules; not the complete cloud memory service | Yes for the app and CLI with native identity; cloud excluded |
| DeepSeek Harness | Versioned local session logs | Instruction files | No |
| Devin | CLI history, Desktop events, explicit cloud import, cloud session browser capture | Cloud Knowledge API, with a Devin API credential | Yes for CLI; no separate Desktop/cloud adapter |
| Factory Droid | Local history and explicit cloud sessions | Project files only | Yes locally with native identity |
| Gemini web | Takeout export or browser capture | No saved-memory reader | No |
| Gemini CLI | Historical local history | Global and project GEMINI.md |
Historical import only |
| GitHub Copilot | Local CLI / IDE history; explicit coding-agent cloud import | Project files only | Yes for CLI; explicit binding |
| Goose | Local JSONL and SQLite histories | Global/project hints and native memory files | No |
| Grok Build | Local session history | Global, workspace, and session-log memories | Yes |
| Hermes | Profile SQLite history and fallback logs | Profile memory, persona, and workspace instructions | No |
| iFlow | Historical local history | Project files only | No |
| Jules | Explicit cloud import | No dedicated memory reader | No |
| Junie | Local CLI and IDE history | Guidelines, agents, commands, project memories and backups | No |
| Kilo Code | Current SQLite, legacy tasks, explicit cloud import | Rules, agents, workflows, project memory files | No |
| Kimi Code and Kimi CLI | Local histories | Project files only | Yes; explicit binding |
| Kiro | Unified, CLI, IDE, Crew history and native exports | Steering, specs, powers, agent prompts, Crew memory records | Yes; explicit binding; agent-specific overrides may apply |
| Meta Muse Code | Local history | Project files only | No |
| MiMo Code | Local SQLite history | Project files only | No |
| MiniMax Code | Shared CLI/app history and explicit cloud import | User/agent Markdown memories, project instructions | No |
| Mistral Vibe | Local monolithic and split histories | Project files only | No |
| OpenClaw | Gateway profile SQLite and retained histories | Workspace instructions, daily memories | No |
| OpenCode | Local JSON and SQLite histories | Project files only; can retrieve other providers’ archived memories | Yes |
| OpenHands | Local CLI/SDK/server/Canvas and legacy histories; explicit cloud import | Local memories, always-on repository microagent, project instructions | No |
| Perplexity | Browser capture, with research steps and cited sources | No dedicated memory reader | No |
| pi | Local history | Project files only | No |
| Prime Agent | Local history | Global continual-harness state and refinement history; session-local state stays with the session | No |
| Qoder | CLI/app/IDE/Work history and explicit cloud import | Rules, auto memory, Work awareness, IDE records; referenced cloud memories during cloud import | No |
| Qwen Code | Historical local history | Project files only | No |
| Roomote | Explicit database history copy or cloud transcript import | Database import captures retained memory outbox text and deployment instructions; periodic backup does not query Roomote | No |
| Roo Code | Historical IDE and CLI tasks | Rules, commands, custom modes | No |
| TRAE | Agent, legacy IDE, app, and Work histories | Local rules, knowledge Markdown, read-only core database memories | No |
| Warp | Local desktop/CLI history and explicit cloud import | Instructions, cached Warp Drive rules | No |
| Windsurf | Historical/local readable history, exports, explicit recovery | Global rules and local memory records; binary records are not searchable as text | No |
| Zed | Native Agent Panel SQLite and historical text threads | Instructions, historical Rules Library | Yes; explicit binding |
All successfully archived conversations share the archive’s search, viewer, export, and sync facilities. Available tools, models, token usage, attachments, and edits depend on what each source actually records. For example, Codex cloud tasks read through the Codex CLI alone lack a full transcript, Jules lacks model/token usage, and binary Antigravity history is not decoded into tool steps. CodeBuddy’s closed-source format has limited real-store verification. Consult the individual source guides for these limitations.
Local memory backup runs every 15 minutes while the watcher is running. Cloud conversation imports and account exports are explicit operations. Devin knowledge capture is the credential-dependent API reader inside memory backup; Claude.ai, ChatGPT, Qoder cloud, and Roomote have the intake paths shown above. A missing provider directory or credential does not produce a complete memory archive for that provider.
Cross-provider memory retrieval
Section titled “Cross-provider memory retrieval”An agent with search_memories can find memories from every provider present
in the current Hansard archive, including records received through Personal
Sync. A Codex conversation can read Claude Code memory, and a Claude Code
conversation can read Codex memory. The calling model does not restrict the
memory provider.
| Behavior | Current support |
|---|---|
| Search across providers | Yes; omit the optional provider filter |
| Search across projects | Yes; omit the optional repo filter |
| Search one project plus global guidance | Supply repo; matching project results come before global results |
| Exclude global guidance | Set include_global: false with repo |
| Fetch matching content | get_memory reads a window of the current snapshot, up to 50,000 characters |
| Search content and metadata | Yes, using local full-text search and a direct-scan fallback |
| Search every historical revision | No; search and MCP fetch use the current snapshot |
| Browse older revisions | Yes, in the Memories view |
| Read binary memory content as text | No; metadata remains searchable |
| Automatically load the whole memory bank into every conversation | No; the agent retrieves relevant records when needed |
| Automatically combine contradictory memories | No; unify-memories helps review and apply selected changes |
| Write through memory MCP tools | No; both tools are read-only |
| Edit a local provider memory | The Memories view can edit a writable local file; cloud, synced, and transcript-only records are read-only |
| Fetch private cloud memories that were never imported | No |
Project and provider filters select search results; they are not isolation boundaries between agents sharing the same local archive. Retrieval reads captured snapshots, so a provider file changed after the latest backup can be newer than its archived copy. Missing provider originals do not remove readable archived snapshots.
Memory retrieval is included in ordinary agent integrations. To refresh one:
hansard integrations add-to codexReplace codex with claude, grok, antigravity, devin, or opencode and
reload the client. hansard integrations sync upgrades older recorded installs
and preserves explicit --no-memories choices. No archive reimport is needed.
To inspect the bank without changing integration settings:
hansard memory list --jsonhansard memory search "test conventions" --limit 10 --jsonhansard integrations status --jsonAn installed integration, an existing memory archive, and tools attached to a running conversation are three separate checks. A healthy memory archive does not imply that an already-open MCP connection exposes memory tools.
See Memories, agent integrations, continuation, agent messaging, and accounts and routing for setup and use.