handover.tools
Server Details
Hand work between AI agents across devices and vendors: one-shot links or an E2EE channel.
- Status
- Healthy
- Uptime
- 100.0% over the last 24 hours
- Last Tested
- Transport
- Streamable HTTP · MCP 2025-11-25
- URL
- Repository
- gammaland/handover
- GitHub Stars
- 0
- Server Listing
- handover.tools
- ChatGPT
- –
- Claude
- –
TDQS
Scored across 9 tools
Each tool maps to a distinct protocol step or handoff operation: channel_* tools separate open, join, send, fetch, ack, and close; handoff_* tools separate put, get, and revoke. The verbose descriptions clarify boundaries even where fetch/ack are tightly coupled.
Consistent snake_case namespace_action pattern throughout: channel_ prefix for channel operations and handoff_ prefix for one-time handoff operations. Verbs are clear and no conventions are mixed.
Nine tools are well-scoped for two related workflows: six for the channel lifecycle and three for one-time handoff. Each tool has a clear role, and there is no redundant or unnecessary tool.
The surface covers the full lifecycle for both workflows: open, join, send, fetch, acknowledge, and close for channels; store, retrieve, and revoke for handoffs. No obvious missing operations for the stated secure handover purpose.
Publisher details
- Operator
- GammaLand · Publisher source
- Operator website
- https://handover.tools
- Vendor relationship
- First-party · Publisher source
- Documentation
- https://handover.tools
- Privacy policy
- Not provided
- Trust center
- https://handover.tools/security
- Restrictions
- None for setup: free, no account or sign-up, no OAuth, no regional limits. Calls are rate-limited per caller per day.
Available Tools
9 toolschannel_ackAcknowledge the paired agent's messages after processing themAInspect
Use the local client at https://handover.tools/client/channel.py rather than calling this by hand. After a fetch with peek=true, the client calls this with the highest sequence number it processed. That is what the peer sees as read, so a response lost in transit never turns into a false read receipt. The cursor never moves back or past the last stored message; calling again with the same seq changes nothing. Slides the channel's expiry.
Inputs: key (required), seq (required).
channel_closeClose a channel and delete its messagesAInspect
Either side can close. Deletes the channel, both keys, and every stored message immediately. The other side's next fetch reports not_found_or_closed.
Inputs: key (required).
channel_fetchFetch unread encrypted messages from the paired agentAInspect
Use the local client at https://handover.tools/client/channel.py rather than calling this by hand: it returns ciphertext only the client can decrypt. Returns the peer's messages this side has not fetched yet. Call only when the person asks. Clients fetch with peek=true, which marks nothing, and then call channel_ack once they have processed the messages; without peek, the messages are marked read when the server sends them (kept for older clients); after=N returns messages after sequence N (the client passes the last one it processed, so a response lost in transit is fetched again; a smaller N re-reads). Side A also receives the peer's SPAKE2 message here to finish pairing. Large backlogs are paged: more=true means fetch again.
Inputs: key (required), peek, after.
channel_joinJoin a channel with the code from the other machineAInspect
Second half of pairing, run by the local client at https://handover.tools/client/channel.py. Send only the first 3 characters of the spoken code (the nameplate) plus this side's SPAKE2 message; the other 5 characters are the password and must never be sent to the server. The nameplate is consumed on success. Returns this side's key and the peer's SPAKE2 message. Unknown, used, or expired nameplates are deliberately indistinguishable.
Inputs: pake_msg (required), nameplate (required).
channel_openOpen an end-to-end encrypted channel to an agent on another machineAInspect
Start pairing two agents on different machines for a multi-round exchange (e.g. one explores, the other studies and asks follow-up questions). Use the local client at https://handover.tools/client/channel.py — do not call this by hand. The client generates a secret password locally, and this call returns only a 3-character nameplate; the person gives nameplate + password (8 letters) to the other machine. The server never sees the password, so it cannot impersonate either side (SPAKE2). The person stays in the loop: send only after they approve, fetch only when they ask. Counts as one write against the daily write quota.
Inputs: pake_msg (required).
channel_sendSend one encrypted message to the paired agentAInspect
Relay one ciphertext to the other side of a paired channel. The ciphertext must come from the local client at https://handover.tools/client/channel.py; do not call this by hand, and never base64-encode plaintext into it (that is rejected). Only send after the person has seen and approved the draft — the server cannot enforce this, the calling agent must. Returns the message's sequence number and how far the peer has read.
Inputs: key (required), ciphertext (required).
handoff_getFetch context handed off from another deviceAInspect
Fetch content stored by handoff_put using its 8-character code. The code is case-insensitive and hyphens are ignored, so 'kvmtrhxp' and 'KVMT-RHXP' are the same; the person may give it as a link like handover.tools/kvmtrhxp, and passing that is fine too. Return the content to the person verbatim and complete: it can be fetched only once. If the code is unknown, already fetched, or expired, content is null with a reason — these cases are deliberately indistinguishable. A channel link (handover.tools/join/, or anything the person said with "Join") is not fetched: join it with the channel client (see /llms.txt). Passed here, it returns reason channel_code with the steps.
Inputs: code (required).
handoff_putHand off context to another deviceAInspect
Store a piece of context and get back a short code. The person gives the link handover.tools/ to an agent on another device and it can fetch the content — no copy-paste of long text. The code is 8 letters, easy to type on a phone. NOT ENCRYPTED: the server can read what you store, so never put credentials, tokens, or private data in it, and ask the person before storing an employer's or client's confidential material (security findings, internal system details, unreleased code). Content becomes unreadable after the TTL expires, and by default it can only be fetched once. The response also returns a revoke_key — keep it for the rest of this session so you can destroy the content immediately with handoff_revoke if the person changes their mind. Writes are limited to a small number per day; reads are not.
Inputs: content (required), label, ttl_minutes, burn_after_read.
handoff_revokeDestroy handed-off content immediatelyAInspect
Destroy content stored by handoff_put right now, before its TTL runs out and whether or not it has been fetched. Requires the revoke_key returned by handoff_put — the person fetching with the short code cannot revoke, only the writer can. Use this the moment someone says they sent the wrong thing, or changed their mind. Revoking an already gone item is not an error: the outcome is reported as 'already_gone'.
Inputs: revoke_key (required).
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
Related MCP Connectors
Agent to agent calls across AI apps. Humans remain in authority.
End-to-end encrypted messaging and work coordination for autonomous AI agents.
Cross-device AI memory with encrypted activity capture and context handoff between AI tools
Matchmaking network for personal AI agents: private agent-to-agent compatibility rendezvous.
Related MCP Servers
- AlicenseAqualityBmaintenanceCross-device communication between AI agents over the public internet231MIT
- AlicenseNot gradedqualityBmaintenanceOpen, vendor-neutral channel for AI agents: signed identity, offline delivery, human approval.3MIT
- AlicenseNot gradedqualityAmaintenanceEnd-to-end encrypted, real-time messaging between AI agent sessions across machines, accounts and vendors.6Apache 2.0
- AlicenseNot gradedqualityBmaintenanceEnables desktop AI agents in different apps to hand tasks to a bound worker chat on the same computer, with per-task approval and isolated, consent-bound replies returning to the initiating chat.1MIT
Glama MCP Gateway
Add one secure layer between your agents and this server.