Recondo sits on the wire between coding agents — Claude Code, Codex, Cursor, Aider, Gemini — and their LLM providers. Every request and response flows through it. Visible in one place. Controllable at the byte level.
On-prem gateway. Multi-vendor. Zero touch.
Not an SDK. Not a framework. A network-layer chokepoint that sees every byte your coding agents send, and gives you the wire to control it.
Sit on the wire between agents and providers. No SDK swap, no agent code modification, no developer workflow change. Trust the gateway CA and traffic flows through.
Anthropic, OpenAI / Codex, and Gemini through one gateway. One dashboard, one set of controls, one invoice reconciliation across the whole agent fleet.
Trace every request to a developer, repo, or team. Cap spend per team before the next runaway loop closes the books. The $380K invoice now has names attached.
Redact secrets, route by model, throttle by team, or block a provider — at the network layer, before bytes leave your network. Control where it actually matters.
Every request and response is SHA-256 hashed and stored content-addressed.
recondo verify re-hashes and compares. Bytes don't change silently after the fact.
Keep agent data in your environment. Local filesystem for dev, your own S3 + KMS in prod. Object Lock COMPLIANCE mode, customer-managed keys.
Agents CONNECT through the gateway. TLS terminates inside.
Bytes are captured, hashed, compressed, parsed, and re-encrypted upstream.
No SDK. No instrumentation.
Per-host leaf cert generated on the fly. Bytes captured both directions.
SHA-256 content addressing. Gzip on disk. Independently re-hashable.
Provider-aware parsers extract tokens, model, cost, tool calls.
Append-only sessions, turns, tool calls. UPDATE/DELETE blocked at the trigger.
A live feed, a session timeline, hash-verified audit trails, cost intelligence, and agent analytics — all reading from the same captured byte stream.
A live feed of every captured request across providers. User turns per minute, active sessions, tokens and cost in the last hour, P95 latency — all updating as new turns close. Filter by Anthropic, OpenAI, or Gemini and the table rewinds to that lane.
Filter by framework — Claude Code, Cursor, Codex, Aider — or by status. Sort by tokens, cost, duration. Drill into any session for the full turn-by-turn trace.
Each turn shows the prompt, input/output tokens, cache reads vs creates, request and response SHA-256 hashes, TTFB, stop reason, transport, and capture status. Click a turn — full prompt, full response, full tool call list.
When an agent uploads an image — a screenshot, a diagram, a UI mock — Recondo captures the multipart body, hashes it, and the dashboard renders the original image alongside the model's analysis. The audit trail isn't text-only; the bytes the model actually saw are preserved and re-displayable.
Every turn, every byte, SHA-256 hashed at capture and re-verifiable on demand. Export to CSV for the auditor.
Spend by provider, model, framework, day, developer. Cache hit rate. Three-month projection with named assumptions.
Which agents are active, which developers run them, how the framework split is shifting. Favorite model per developer.
Self-host the dashboard alongside the gateway with just dashboard-dev.
The same capture stream feeds two more surfaces. Same database, same project scoping, same auth model — pick the one that fits your terminal.
Live realtime lens in under 500 ms. Five metric cards on top, a 50-row virtualized traffic table underneath. Sessions, cost, agents, audit — vim navigation, command palette, pinned tabs, selection that follows you across lenses.
$ cargo run -p recondo-tui
d realtime · s sessions · c cost · a agents
:today · :week · :since 2026-04-01
A long-running MCP service at :4001/mcp. Claude Code, Cursor,
Goose, and any MCP-capable agent gain 28 read tools and 7 action-gated tools
over their own captured history — sessions, turns, cost,
anomalies, compliance, governance.
$ just fullstack
$ recondo-mcp config claude-code \
>> ~/.claude/mcp_servers.json
# Then restart Claude Code; /mcp shows "recondo" registered.Three audiences, one gateway. The same captured byte stream serves platform leaders, finance, and security.
Real coding-agent usage across teams, providers, and projects — not what dashboards from three different vendors say about themselves.
The Anthropic invoice is one number. Recondo turns it into a thousand — one per developer, repo, branch, or team — so spend caps actually work.
Redact at the wire. Block providers per environment. Append-only capture and S3 Object Lock turn audit into receipts, not vibes.
"Deleted from the database" is not the same as "deleted from the object store." Recondo's capture pipeline is engineered so what you captured is what stays captured.
turns and tool_calls are write-once. A PostgreSQL trigger refuses any modification — captures cannot be silently rewritten.
Every body is hashed and stored by hash. recondo verify <session-id> re-hashes the bytes and compares.
Bodies live in a versioned S3 bucket with Object Lock COMPLIANCE retention (365 days default). Even root can't delete.
Server-side encryption with your own KMS key, automatic rotation enabled. The cloud provider can't read your captures.
Agent ↔ gateway is TLS. Gateway ↔ provider is TLS. The plaintext exists only inside the gateway, only long enough to capture.
Standard → Infrequent Access at 90 days → Glacier at 365 days. Retention without the storage bill.
Build the gateway, generate a CA, route Claude Code through it. That's the whole loop.
just setup
just build
just runListens on :8443. Generates a CA and installs it into the system trust store the first time.
NODE_EXTRA_CA_CERTS=$HOME/.recondo/ca/ca.crt \
HTTPS_PROXY=http://localhost:8443 \
claudeEvery prompt, every tool call, every response — captured, hashed, persisted.
just recondo sessions
just recondo session sess_01HZ…
just recondo verify sess_01HZ…CLI inspector ships with the binary. Re-hash captures any time and compare.
Recondo is a Rust gateway, a TypeScript GraphQL API, and a React dashboard — all in one repository. Run it on your laptop with SQLite + local filesystem. Run it in your VPC with PostgreSQL, S3 Object Lock, and a customer-managed KMS key. Either way, agent traffic never leaves your environment.