Clawpost vs Postfleet

Clawpost 2

The last commit was on 1 April 2026, 219 days before this check (0). No releases or commits in the 90 days from 8 August (0). Issue and pull request activity couldn't be read through GitHub's API that day (0, unchecked). Not in the official MCP registry and no published package (0). Dependencies such as agents ^0.0.74 and the MCP SDK ^1.12.1 date from early 2026, and there's no CI (2 of 10). Total 2.

Postfleet 75

Read for a closed service. The security page's record of 25 October 2026 says the scan moved to claude-haiku-5-5, a model released on 7 October, and the sending docs were updated on 13 October with account-wide sending holds (30, 10 days before 4 November). Dated changes in the 90 days to 4 November include @postfleet/mcp 0.1.2 and 0.1.3 on 24 August and 0.1.4 on 2 September, the SDK 0.1.0 on 8 September, docs updates on 12 and 18 September and 13 October, and the terms of 6 October (20). No public changelog (/changelog still 404) or issue activity we could read. The GitHub mirror's last commits, on 19 and 27 September, carry no version or tag and don't count as releases; GitHub pages beyond the repository front page were refused to our reader. Support is by email to a mailbox the site says an agent reads, untested (5 of 15). Listed in the official MCP registry as ai.postfleet/postfleet, a domain-verified namespace (15). The MCP package has two current dependencies (@modelcontextprotocol/sdk ^1.26.0, zod ^4.4.3) and a publish workflow with npm trusted publishing, but no tests or CI build, and 0.1.4 was published by hand without provenance (5 of 10).

Clawpost 60

Free self-hosted software with nothing to buy from the author, which the rubric scores 20 + 20 + 20. Running it needs the deployer's own Cloudflare Workers Paid plan at $5 a month, which isn't Clawpost's charge. No payment protocol (0 of 40). Total 60.

Postfleet 30

No x402, MPP or L402 in the docs or the OpenAPI file (0 of 40). Four plans with public prices and limits, but no per-message metered price (10). Free has no card, with 100 comprehended messages, 3 mailboxes and 3,000 sends a month (20). A person has to sign up and create keys in the dashboard. The homepage says the agent provisions its own address with no account page, which is true of mailboxes once a key exists, but not of the key. The demo inbox needs no signup but is a demo under /api/demo, not API access (0 of 20).

Clawpost 54

All 21 MCP tools take typed zod inputs, which the MCP SDK publishes as JSON Schema (25). SKILL.md and AGENTS.md give agent-readable setup in Markdown; no llms.txt (5 of 10). Descriptions are one line each, such as Send an email, with parameter descriptions but no when-not-to-use (8 of 20). Inputs are typed with defaults for limit and offset, but recipients are a comma-separated string and direction isn't an enum in the schema read (9 of 15). The README has request examples and a REST route table; error responses aren't documented (7 of 15). No versioning or changelog (0). Total 54.

Postfleet 83

OpenAPI 3.1 at postfleet.ai/openapi.json with 10 paths and 15 operations, closed request and response schemas, enums for statuses and authentication verdicts and a typed schema per error code. comprehension and attachments on a message are open objects, and the servers entry is a relative path with an example origin (22 of 25). llms.txt, llms-full.txt and every docs page as Markdown at /docs/<slug>/raw (10). The 16 MCP tool descriptions in the published package say when to use each, when to use another tool instead and what not to retry, and the docs pages each end with a current limitations section (18 of 20). Inputs are typed with UUID and email formats, length limits, a client_id pattern and bounded limit and timeout_seconds; a few strings are free text (13 of 15). curl, TypeScript and Node examples, a replay table and named error codes with their status (key_scope, send_blocked_by_list, idempotency_conflict, sending_held, delivery_outcome_unknown) (15). The API is at /api/v1 and docs pages carry updatedAt dates, but there's no changelog; /changelog answers 404 (5 of 15). Most curl examples in the docs point at https://your-app.example.com rather than api.postfleet.ai.

Clawpost 8

Graded with the lines for self-hosted software, since each deployer runs their own Worker. No published package; it deploys from the repository with a Deploy to Cloudflare button, Bun and Wrangler, and states the Cloudflare products it needs (8 of 20). No CI workflow or test suite in the repository (0). Open issues couldn't be read, because GitHub's API refused our requests for its rate limit, so the line scores absent (0, unchecked). No tags, releases or changelog; package.json says 0.1.0 (0). Pre-1.0 (0). Total 8.

Postfleet 35

Graded with the hosted lines, on the REST API with the MCP server noted. No status page found. /status still answered 404 on 4 November and none is linked (0). No readable incident history (5). Sending limits are published with numbers (holds at 20 attempts a minute and 100 an hour, Free at 100 sends a day and 3,000 a month, monthly quotas per plan), but read limits aren't, Free's new-recipient limit has no number and the comprehend route's hourly cap is unstated (10 of 15). Sends take a client_id with a full replay table, errors carry retryable, and the docs say which outcomes to retry with the same key and which need a new one. API 429s document no Retry-After or reset time (10 of 15). No SLA found on any plan (0). Nothing on the site calls the API or MCP server beta or preview, though both packages are 0.x (10).

Clawpost 42

One shared secret for REST and MCP, compared in constant time, with no scopes, expiry or per-client keys (10 of 30). Sender approval, drafts, and an optional allowlist of domains for routing changes limit what an agent can see and change; the key can still send to anyone (14 of 20). Mail from unapproved senders is hidden from all query tools by default, and list_pending returns metadata without bodies, documented as a prompt-injection guard (15). No log of API or tool calls beyond the stored messages (3 of 15). No security policy, advisories or disclosure route for this personal project (0). Total 42.

Postfleet 68

Bearer pf_ keys stored hashed, each with one scope (bootstrap, api or mcp), optional binding to one mailbox and can_read and can_send flags, revocable in the dashboard and sent in the header only. No rotation flow is described beyond issuing a new key, and the MCP OAuth flow attaches an existing key rather than issuing a scoped token (25 of 30). A key can be read-only and bound to one mailbox, cross-scope IDs return 404, send lists and approval settings are dashboard-only, and require_approval holds sends for a person who approves a hash of the exact text (20). For the prompt-injection check the docs set out sanitisation with a removal report, a deterministic pre-scan and a model scan that fails closed, SPF, DKIM and DMARC verdicts on every message, and MCP reads without the raw body. We haven't tested any of it, and the 99.1% catch rate is the vendor's claim on a private corpus. REST message reads still include body_raw for seven days, and flagged mail reaches the agent as cleaned text with a flag (14 of 15). Approvals and rejections are audited, draft create, edit and delete aren't (the docs say so), and no per-call log was found (6 of 15). No security.txt on postfleet.ai (404 on 4 November) or api.postfleet.ai, no disclosure policy or bug bounty, and the privacy policy says it doesn't claim SOC 2. The published red-team record and npm trusted publishing on two of five package versions earn a little (3 of 20).

Clawpost 53

21 tools, the 11 to 30 band (15 of 25). List and search tools take limit, offset and filters by direction, sender, label and archive state (20). Error shapes aren't documented; the code returns JSON errors (8 of 20). No idempotency key for sends and no readOnlyHint or destructiveHint annotations; drafts give a review step (3 of 20). Archived mail is excluded by default and the CLI has a --json mode, but there's no SDK beyond the CLI (7 of 15). Total 53.

Postfleet 72

Graded on the REST API and the MCP server together. Sixteen MCP tools with descriptions of two to four sentences, and no toolsets or dynamic loading; the vendor's tutorials show client-side allowlists of three read tools. Message lists return metadata and comprehension status without bodies (17 of 25). Listing takes only limit (1 to 100), with no cursor, date filter, thread filter or search; wait_for_email filters by sender and subject substrings (8 of 20). Errors carry a stable code, retryable and, on MCP, a recovery hint per code and status (19 of 20). client_id is required on MCP sends and optional on REST, replays are specified per state, and every tool carries readOnly, destructive and idempotent annotations in the package source (20). Few required fields; one official SDK, TypeScript 0.1.0, and no Python SDK, which the vendor confirms (8 of 15).