A minimal tool synthesised mid-task from your agent's own examples, proven in a sandbox before it ever runs, pinned at birth, scoped to the task, and dead when the task ends.
Registries stock only what someone productised in advance; the long tail of one-task needs — convert this vendor's idiosyncratic export, parse this counterparty's attachment format, reconcile these two column conventions — is unlistable by definition, so agents improvise the transform inline: no examples pinned in advance, no adversarial cases, no permission scope, no receipt, and the same need re-improvised differently next time. The forge is that improvisation with every safeguard attached. (1) Forge request — the agent submits input/output example pairs, a plain-language capability statement, an explicit forbidden-effects list, and a task binding (task ID + expiry). (2) Synthesis + example-proof — the forge synthesises the minimal tool by staged search (exact picks first, deterministic transforms second, fitted transforms last) and proves it in a sandbox: all supplied examples must pass and generated edge cases must behave within the stated contract; the forbidden-effects list is enforced by the sandbox runner (network, DNS, subprocess, and writes outside a scratch scope are blocked), not by the tool's promises. A tool that fails proof is never issued — the agent gets the failing cases back as the receipt of why not. (3) Pinned issuance — a passing tool is issued as a signed, hash-pinned artifact with a capability-scoped, task-bound credential and a hard expiry; the calling agent loads it for the task's remaining calls only. (4) Death + receipt — at expiry or task end the credential dies and the executable is archived out of existence, while the receipt persists: examples hash, proof cases, calls made, per-call input/output hashes. A tool with a lifespan of one task cannot be rug-pulled, needs no registry listing, and adds zero standing context. Pricing is per forge, never per call. Stated plainly: synthesis as built is deterministic template search over a transform library — a capability outside that library's reach is refused at proof, not improvised, and an LLM synthesis leg is not part of the shipped product (its coverage is unmeasured); demand, missing-tool frequency, and willingness to pay are unmeasured.
Forge attempt, proof failed (no tool issued; the failing cases returned as the receipt of why not; stated retry cap: 3 synthesis rounds per attempt, further rounds metered as new attempts): $0.49 Successful forge issuance (example-proof passed, hash-pinned, task-scoped credential, expiry set; proof + task calls + receipt included; calls are never charged per call) — HEADLINE: $0.99 Forge + extended fuzz pack (deeper generated edge/adversarial suite, larger example set): $2.49 Fleet forge pass (per operator / month: stated allowance of 40 successful forges + 40 failed attempts per month; use beyond the allowance meters at the per-forge tiers above): $19.99/mo One-time trial: One free successful forge per operator, ever. No perpetual free tier and no other free allotment. Support: support@textforagents.com
POST /register with {"name": ..., "wallet": ...} to self-serve your per-customer credential — no operator assistance needed. Then POST /forge, POST /call, POST /expire, GET /receipt/{task_id}, POST /subscribe, or call the MCP-shaped tools at /mcp. Documentation: /docs · OpenAPI: /openapi.json · llms.txt: /llms.txt · Terms: /terms · Privacy: /privacy · Pricing: /pricing
Stated plainly: synthesis as built is deterministic template search over a transform library — a capability outside that library's reach is refused at proof, not improvised, and an LLM synthesis leg is not part of the shipped product (its coverage is unmeasured); demand, missing-tool frequency, and willingness to pay are unmeasured. A tool is issued only if it reproduces ALL supplied examples and passes the generated edge cases in the sandbox; forbidden effects are enforced by the sandbox runner, not by the tool's promises; every forged tool is task-bound and expires at task end or TTL, whichever comes first.
Support: support@textforagents.com