coder.
Decision and code intelligence for AI agents.
Record every choice. Map every symbol. Surface both at the right time — as context, fine-tuning data, RL signal, or replay. Local-first. Host-neutral.
Every token your agent burns figuring out what's around it is a token it can't burn on the actual problem.
Most agent failures trace to context. Too little, and it guesses. Too much, and it drowns. Forge fixes both directions: blast radius scopes the change to its real callers, sidecar-first projections replace raw source dumps with structured maps, and just-in-time hook injection surfaces the right decision on the turn it matters — not jammed into every prompt.
Project-scoped recall keeps other codebases out of this one. The ubiquitous language of your repo is in the agent's hands by the second turn.
100 memories. 60 seconds.
Install. Bootstrap. Your agent remembers everything.
Reconstruct every choice your agent made.
Forge captures the decision trajectory — what the agent considered, what it picked, what evidence it cited, what it deferred. Local-first, replayable, yours.
Every choice in order, with evidence, counter-evidence, and a path-score.
Branch points, dead ends, the path actually taken vs the paths considered.
Replay any decision with different evidence; see what the agent would do.
Every suggestion grounded in your real code.
Forge maintains a live map of the codebase — symbols, imports, references, tests — that any agent can query before suggesting changes. No hallucinated symbols, no broken refactors.
find_symbol, references, imports, callers — instantly, scoped to the active project.
Code intelligence is read from local sidecar projections (CodeMap, GraphShard). No cloud upload, no API key, no hosted index.
See every caller and every linked decision before you change a function.
Underneath: 8-layer Manas memory. Above: decision intelligence and code intelligence.
The memory grooms itself. Consolidator phases dedup near-duplicates, decay irrelevant memories, link related ones, and detect contradictions — so what gets recalled is curated, not just stored.
Spawn a team. Watch the team's decision graph. Replay the team. Forge tracks parent-child agent relationships across hosts so multi-agent runs are inspectable end-to-end.
Inspect what your agents are doing.
Forge Studio is the web-first control plane served by the daemon over HTTP/SSE. See live decision timelines, code graphs, memory layers, and agent sessions across hosts. Local-first, no cloud auth.
Live replay of every step, with scoring + evidence inline.
Navigate symbols, references, and impacted callers with one click.
See Claude Code, Codex, Cline, Cursor, Hermes activity unified in one pane.
Your data. Your machine. One file.
No account. No cloud. No trust required.
Single SQLite file on YOUR disk
No cloud. No telemetry. No account for free tier.
You own the file. Delete it. Move it. Back it up.
Bring your own model for air-gapped operation.
Your agents come and go. Forge stays.
One daemon serves Claude Code, Codex, Cline, Cursor, Gemini CLI, and Hermes — the same memory and decisions are visible to all of them. Switch agents mid-task and the next one picks up where the last one left off.
Models change every quarter. Your work doesn't restart.
Smarter agents, every loop.
Every trajectory your agent produces is structured data. Use it.
Decision and code intelligence are the substrate. Use them however you want.
Recall + decision graph in the next prompt.
Every suggestion grounded in your real symbols.
Exportable, scored, privacy-safe trajectories for fine-tuning, RL, distillation, or search-based optimization.
The loop closes itself.
TOOLING FOR LLMS
- Tool-routing models — pick the right tool sequence per (model, language, intent) from your trace data
- Context compressors — predict what the agent will cite, prune the rest
- Self-improving prompts — system-prompt gaps surface from failed-trajectory analysis
- Eval generation — every successful trajectory becomes a labeled eval case
More in /docs#tooling-for-llms.