Roadmap

​

Live today

• Any surface → a verified graph. OpenAPI, human docs, llms.txt, Anchor IDLs, raw program source. One command: gecko add <spec-or-docs>.
• Provenance on every edge: extracted / recovered / flagged on-chain; EXTRACTED → DECLARED → INFERRED → CLAIMED → VERIFIED / REFUTED on HTTP.
• find_start: intent → the right starting point, with a dependency-ordered derive plan and an honest no-start below the retrieval floor. See find_start.
• simulate → Receipt: a simulation before money moves, live-proven on a fork and on mainnet. See The Receipt.
• The binding. prepare_purchase, prepare_instruction and plan_swap return a binding over the exact simulated bytes at exact strength. verify_signed_transaction checks the signed bytes against it, and submit_transaction refuses to relay anything that does not verify. 50 landed mainnet transactions in the ledger as of 2026-09-01.
• Hosted receipts. The orquestra surface at mcp.geckovision.tech/orquestra/mcp simulates against a public mainnet RPC and returns the receipt and the binding; no local install.
• Invisible auth: the credential lives in your OS keychain and is injected at call time. Never in a tool def, never in mcp.json, never in the model’s context.
• $0 recorded mode: falsify every call offline, from the schema, before you spend a token.
• The Scorecard: gecko report <spec>, a grade plus the specific findings.
• Correctness in CI: gecko test <spec>.
• Seven fail-closed security layers, including an AST-enforced no-sign boundary on the landing layer.
• The categorical corpus + an N-confirmed drift detector: gecko drift <series>.
• Hosted MCP: one-click add to Claude Code / Cursor / VS Code. Every client’s wiring is in mcp-config.json.

​

Building

• The drift scheduler. The detector works and the series reads back; we are building the cadence that re-simulates on a schedule so the series accrues without you running anything. Until then it accrues only when runs happen.
• Catalog breadth. 4,500 projects are listed and seven program configs ship (Pump.fun, Meteora, Jupiter, ORE, MetaDAO, Orca Whirlpool, let_me_buy); we are widening the auto-comprehend path so more of the catalog resolves without hand-authoring. Non-Anchor generalization is proven once.
• The TEE credential backend.

​

Exploring

• A semantic/vector retrieval tier. Deliberately OFF behind an evidence gate. It flips only on measured lexical recall failure, not fashion, and one measured negative result on embeddings is already on record.
• Cross-customer episodic pooling. Tenancy is local-only until a consent and egress layer exists that we’d be willing to defend.
• Response-level trust. “Is this response sane?”, deliberately downstream of comprehension and verification.

​

What we’ll never build

• No signing, no custody. Gecko holds no key. Every money path ends in unsigned, checked bytes plus a receipt with a binding; somebody else signs. The one send path, submit_transaction, relays only signed bytes that verify against that binding. Unsigned assembly for simulation is the one documented carve-out, and the landing layer’s no-sign boundary is AST-enforced.
• No payment rail, no marketplace. We compose on MCP, x402, and payment catalogs; the provider charges you directly.
• No storage of your API responses, request values, or secrets. The corpus is categorical by construction: a fail-closed allowlist, not a policy you have to trust.
• No browsable provider marketplace. Point us at the surface; the flow is identical for every API.

​

Who pays

​

Tell us what to comprehend next